AI Agent Frameworks · independent, you own the build

AI agent frameworks in 2026: which to use, and why it's a decision you live with

Pick the wrong one and you don't just swap a library. You rewrite your orchestration, your state handling and your evaluation harness — usually six months in, usually under deadline. This page gives you the current map, the part of the stack that actually decides whether you ship, and an honest answer on whether you should build it yourself at all.

Framework-independent adviceBuilt on open frameworks$8K–$35K — you own it
Heads up, 3 April 2026: Microsoft shipped Agent Framework 1.0 and moved Semantic Kernel and AutoGen into maintenance mode. Most "best frameworks 2026" lists still tell you to choose between them.
US-registered LLC · you own the build
Tell us the workflow — we'll name the frameworkFree, independent, no obligation
We build on open frameworks and hand over the code — no resale, no lock-in. If your project doesn't need a custom build, we'll say so.
88%of agent pilots never reach production (Forrester / Anaconda, 2026)
9,400+public MCP servers — tool connectivity is now a standard
22%of production deployments coordinate three or more agents
100%owned by you — code, prompts, orchestration & evals
Quick answer · what is an AI agent framework?

An AI agent framework is the code layer that decides how your agent thinks and acts: how steps are ordered, how state survives a crash, how tools get called, and when a human takes over. LangGraph, CrewAI and Microsoft Agent Framework are the common picks in 2026. The framework shapes your architecture — it does not, on its own, make the agent reliable.

How LoopHawk helps: you should not have to bet a quarter on a framework guess. We choose it with you, build the agent on your own stack, and show it working on your data before you pay anything. Talk to us →

Are you facing this?

Which of these is actually your problem?

Six ways a framework decision goes wrong. Find yours — the fix is different for each, and none of them is “read another comparison post”.

Recognize yours in that list? Good — every one of them has a name and a fix, and none of the fixes is “go read another comparison post”.

“We picked one and now we’re stuck”

Six months in, the framework fights the workflow instead of carrying it.

The fix: pull your logic out of its classes first — that alone makes the move cheap. See what a rewrite costs.

“The prototype works, production doesn’t”

It demos beautifully and falls over the moment real traffic and real edge cases arrive.

The fix: almost always the missing evaluation layer, not the framework. That is the 64% blocker.

“Nobody can tell us which one to use”

Every comparison post recommends whatever the author sells, and half are out of date.

The fix: four questions, in order. See how to choose in ten minutes.

“We’re on AutoGen or Semantic Kernel”

Both went quiet in April and nobody told you.

The fix: stop new work there, plan the move on your schedule. See the migration path.

“It works but we can’t prove it”

Someone asks how you know the agent is right, and there is no good answer.

The fix: evaluation and observability as a line item, not an afterthought — it is 25% of our build cost for a reason.

“Which model do we even run on it?”

The framework question is settled and the model question is wide open — and the bill scales with it.

The fix: see the model and cost comparison below.

Recognize more than one?Tell us which and we will tell you what we would do — free, and independent of any vendor.
Show me what you’d do →
Under the hood

What is actually inside the AI agent frameworks you are choosing between?

All AI agent frameworks are six parts in a trench coat. Once you can name them, the marketing stops working on you — because you stop comparing brands and start comparing which of the six a framework does well, and which it quietly hands back to you.

AI agent frameworks bundle six moving parts into one package: a reasoning core, a control model, state and memory, a tool surface, human checkpoints and tracing. Five of those six stay swappable later. One does not, and that is the one worth choosing on.

THE LOOP YOUR FRAMEWORK IS ACTUALLY RUNNING

PLAN

Decides the next step from the goal and what it knows

CALL

Reaches into your systems through the tool surface

CHECK

Validates the result before anything is committed

REMEMBER

Writes state forward so step nine still knows step two

Approval gates sit between CHECK and CALL wherever you want them · every step replayable

Illustrative — the loop every agent framework orchestrates, not a specific client’s system.

Swappable

The reasoning core

Where the model thinks. Nearly every framework treats this as pluggable — which is why “which framework” and “which model” are two separate decisions, and why anyone bundling them is selling you something.

Locked in

The control model

Graph, role, or loop — the agent orchestration model. This is the one part you cannot change later without a rewrite, and it is the reason a framework choice outlives the project that made it.

Swappable

State and memory

What the agent still knows on step nine. Thin memory is invisible in a demo and fatal in production, because demos are three steps long.

Swappable

The tool surface

How the agent reaches your systems. Since MCP, this is the part most likely to survive a framework switch — build it as a server, not as glue.

Swappable

Human checkpoints

Where a person approves before money moves. Frameworks differ enormously here and almost nobody scores them on it.

Swappable

Observability

Whether you can see why it did that. The layer teams skip in month one and rebuild in month four, usually during an incident.

Five of the six are replaceable. Pick for the one that is not.

Before you compare brands

AI agent frameworks, SDKs, runtimes or platforms — which are you shopping for?

People shopping for AI agent frameworks use four words as if they mean one thing. They do not. Buying the wrong category is a more expensive mistake than picking the wrong brand inside the right one, because you cannot fix it by switching vendors.

What it isWhat you getWhat you still ownPick it when
LibraryOpen source AI agent framework building blocks, no opinionEverythingYou have strong engineers and a strange problem
FrameworkAn opinion about control flowHosting, evaluation, on-callThe opinion matches your workflow
Runtime / harnessSomewhere for it to actually liveThe agent logic itselfOrchestration is solved but operations are not
PlatformThe whole thing, hostedYour data — and the exit riskSpeed matters more than control

Decide the category first. The brand argument is downstream of it.

Start here — the lists are out of date

What changed in April 2026, and why are the popular lists wrong?

Two frameworks that almost every 2026 comparison still recommends stopped being actively developed. If your shortlist came from one of those posts, it is already stale.

What the lists still say

“Choose between AutoGen and Semantic Kernel”

Presented as two live, competing options you weigh against LangGraph and CrewAI — exactly as they were in 2025.

What is actually true

Both moved to maintenance mode on 3 April 2026

Consolidated into Microsoft Agent Framework 1.0 — .NET and Python, native MCP and A2A, long-term support.

Pick from a stale list and you inherit a framework nobody is steering. That is what you are actually gambling on.

Get the wording right — most write-ups don’t

Not “merged”

Nothing was folded together. A new SDK was built and the capabilities were consolidated into it.

Not “deprecated”

Microsoft has not formally deprecated either project. Your code keeps running.

Maintenance mode

Still there, still working — just no longer where the roadmap goes (InfoWorld reads it the same way).

Why this matters to you and not just to Microsoft: a framework in maintenance mode runs fine on Monday. The problem arrives in month nine, when you need something that only lands in the actively developed SDK — and by then the move is not a library swap. It is your orchestration, your state model and your evals.

Already on one of them? You are not in trouble.

Nothing broke. Production systems on both are still running. The honest advice is simply: stop starting new work there, and move on your schedule rather than during an incident. Here is the order that keeps it cheap.

Freeze new work

Stop adding features on the old runtime. Every feature you add there is a feature you migrate twice.

Lift your logic out

Move business logic out of the framework’s classes into plain functions you own. Usually a week — and worth doing even if you never migrate.

Run both, prove parity

Stand the new framework up beside the old one and run both against the same evaluation set, so you can show behavior did not drift.

Cut over, one at a time

Switch one workflow, watch it, then the next. Never a big-bang weekend.

Skipping step two is what makes step four expensive. Almost every painful migration we read about skipped step two.

Most single-agent moves land inside the range in the cost table below. Larger multi-agent estates take longer, though rarely for the reason people expect: permissions and approvals eat more calendar than code does.

On a framework that just went quiet?We’ll map what a move actually costs you — in weeks, not adjectives.
Map my migration →
What everyone gets wrong

Does the framework you choose decide whether the project succeeds?

Mostly, no. Framework choice decides how painful your second year is. What kills projects in the first year sits one layer above the framework, and almost nobody comparing LangGraph to CrewAI mentions it.

Roughly 88% of agent pilots never graduate to production, according to 2026 Forrester and Anaconda data. Here is what the teams themselves blamed:

64%Evaluation gaps — no reliable way to prove the agent is right
57%Governance friction — can't trace or approve what the agent did
51%Model reliability — behavior drifts and nobody catches it

Read that list again. Not one of those is an orchestration problem. You can build all three on any framework on this page — and skip all three on any framework too, which is exactly what a demo-quality prototype does before it dies in review.

The framework gets the blame. The missing evaluation layer did the killing.

So here is the thing nobody selling you a framework will say: proving the agent is right is the expensive part, and it is the part that decides if you ship. When we quote a build, that layer is a line item, not an afterthought. You will see it in the worked example further down.

The current map

Which AI agent frameworks actually matter in 2026?

Twelve are worth knowing. Most teams will pick from four. Each row says who it is genuinely best for — not which is "best", because that question has no answer without your workflow.

Four names actually matter here. The rest of the list you have been handed is noise, and ignoring it costs you nothing.

FrameworkOrchestration modelBest forStatus
LangGraphGraph — nodes and edges, durable stateStateful production pipelines that need to survive restarts and human approvalActive
CrewAIRole-based crews and tasksFast prototypes of team-shaped work; the lowest barrier to a first resultActive
Microsoft Agent FrameworkUnified workflows, native MCP + A2A.NET shops and Azure-committed enterprises; the successor path from SK/AutoGenGA Apr 2026
OpenAI Agents SDKExplicit handoffs between agentsTeams already committed to the OpenAI ecosystem wanting minimal glueActive
Claude Agent SDKTool-use loop with long contextLong-running research and document work with heavy tool callingActive
Google ADKAgent Development Kit, A2A-nativeGoogle Cloud / Vertex estates and agent-to-agent topologiesActive
Pydantic AIType-safe, validation-firstPython teams who want structured outputs they can actually trustActive
LangChain DeepAgentsPlanning + sub-agents over LangChainLonger autonomous tasks inside an existing LangChain codebaseActive
Strands AgentsModel-driven loopAWS-leaning teams wanting a light, opinionated runtimeActive
Mastra / Vercel AI SDKTypeScript-first, app-embeddedProduct teams shipping agents inside a web app, not a backend serviceActive
Agno / SmolagentsMinimal, code-first loopsSmall scopes and learning; deliberately little abstractionActive
Semantic Kernel · AutoGenSuperseded by Agent Framework 1.0No new builds — migrate on your own scheduleMaintenance

Status verified 11 Aug 2026. We re-check this table monthly, because it has already changed once this year.

Want this narrowed to one row?Describe the workflow and we'll tell you which framework fits — even if the answer is "none, buy a tool".
Narrow it down for me →
The two most teams weigh

LangGraph vs CrewAI — which should you pick?

Pick LangGraph when the work has state worth protecting. Pick CrewAI when the work is a small team of roles passing a task along and you want something running this week.

Neither one is better. One is better for the shape of your work — which makes this a question about your process, not about code.

Choose LangGraph when…

  • the process has real steps that must survive a crash or a restart
  • a human has to approve something in the middle
  • you need to replay exactly what happened, later, under scrutiny
  • branches and loops matter — retries, escalation, conditional paths

The cost: more upfront design. You will think about state before you write much agent code.

Choose CrewAI when…

  • the job splits naturally into roles — researcher, writer, reviewer
  • you want a working prototype in days, not two weeks
  • the run is short and a failure just means running it again
  • you are validating whether the idea is worth building properly

The cost: less control when you outgrow it. Plenty of teams prototype here and rebuild on LangGraph.

And that last sentence is not a criticism. Prototyping in CrewAI and moving to LangGraph is a legitimate, cheap path — as long as you decide it deliberately instead of discovering it in month six.

The numbers, in proportion

Do the benchmark numbers actually change which one you should pick?

Someone finally benchmarked AI agent frameworks instead of asserting. The results are worth reading — and worth putting in proportion, because the spread they found is real but it is not where your money goes.

Mostly no. Measured framework overhead differs by roughly a tenth on the same task, which is a rounding error next to the model tokens your prompts and retrieved context actually consume. Use them to rule out the genuinely heavy option, not to rank the top three AI agent frameworks.

~11%Spread in orchestration token overhead between the leading frameworks on an identical task
1,288Input tokens for the graph-based run in that test
1,432Input tokens for the role-based run — same task, same model
0Benchmarks that measure whether you can debug it in month six

Figures published by AgentMail’s framework benchmark (8 July 2026), which installed and ran each framework on the same task. We did not run this test; we are citing theirs and adding the reading they leave out.

Why an 11% token gap is not your problem

Run the arithmetic on a real workload before you optimize for it. At a million agent turns a month, eleven percent of the orchestration overhead disappears next to the context you are actually paying for. Teams tune the 11% because it is measurable, and leave the far larger number sitting in an unpruned context window.

What no benchmark can measure

No harness can time the thing that kills projects: whether you can tell why the agent did that, six weeks after the engineer who built it moved teams. Install weight is measurable. Debuggability is not — and debuggability is the one that decides.

Measure what you can. Just do not mistake it for what matters.

The selection axis nobody scores

How much should MCP and A2A support matter to your choice?

Enough to be a tie-breaker. MCP standardizes how an agent reaches tools and data; A2A standardizes how agents talk to each other. A framework with neither will cost you adapter code you did not plan for.

Protocols are cheap to add now and expensive to retrofit later. It is a fair argument, and it holds up.

MCP — vertical (agent → tools & data)

The connective layer between your agent and everything it needs to read or do. Adoption has passed 9,400 public servers, with 50+ enterprise partners implementing it through 2026. If your agent touches internal systems, this is the plug you want it to speak.

A2A — horizontal (agent → agent)

How separate agents — sometimes from different vendors — hand work to each other. In production at 150+ organizations as of April 2026. Matters most if you expect to coordinate agents you do not own.

They are complementary, not competing. Microsoft Agent Framework and Google ADK ship both natively; LangGraph and CrewAI integrate MCP well. Best for: if your agent lives inside one company and calls internal tools, MCP is the one to care about. A2A becomes interesting the moment another company's agent enters the workflow.

Honest sizing

Do you actually need a multi-agent system?

Probably not yet. Only about one in five production deployments runs three or more agents together. Most working systems are one well-built agent with good tools, and they are cheaper to run and far easier to debug.

Most teams reach for a multi-agent framework one project too early, then spend the next quarter debugging conversations between robots.

What actually multiplies when you split one agent into three

Not the capability. Multi-agent looks impressive in a demo because the diagram looks impressive. Here is what the diagram does not show you.

Places a failure can start1 → 5
worst offender
Three agents, plus the two handoffs between them. A handoff is a message one agent writes and another has to interpret — and that is a new failure mode, not a free connection.
Decisions you have to explain afterwards1 → 3
Someone asks why the refund was approved. With one agent there is one trace. With three, you are reconstructing a conversation.
Model round-trips per task1 → 3+
Every handoff is another model call. Latency is additive and your user feels every second of it.
Evaluation sets to maintain1 → 3
Each agent needs its own, or you cannot prove any of them is right — and that is the 64% blocker that kills pilots.
What multipliesYour debugging surface

Straight arithmetic on a three-agent design, not a benchmark. Your numbers will differ; the direction will not.

Four gates before you split

Work down them in order. Most teams stop at the first one.

Does one context window hold it?

If you can describe the whole job to a new hire in one paragraph, one agent will do it. Stop here — most of you can.

Different tools and different permissions?

Different prompts is not a reason to split. Different permissions is — when one role must never be able to issue the refund the other one can.

Do separate teams own the pieces?

An org boundary is a legitimate seam. If two teams need to ship without waiting on each other, split along that line, not along a whiteboard.

Still yes? Split one seam.

One seam at a time, at the place that actually hurt. Never straight into the five-box diagram from the conference talk.

If you stopped at gate one, that is a win, not a shortcut. You just saved yourself a quarter of debugging.

Where we land, plainly: one agent to begin with, and we divide it only when the work itself demands it. It is how we build our own agents, and it has kept them simpler than they would otherwise be.
Been told you need a multi-agent system?Send us the workflow — we’ll tell you honestly if one agent does it.
Check my architecture →
The full solution space

Framework, platform, or delivery partner — which route is right for you?

Three ways to end up with a working agent, and they suit very different situations. The framework is only the first of them, and it is the most work.

Sometimes the honest answer is “buy the platform”. Better you hear that from us now than find it out in month four.

RouteBest forWhat it really costsYou end up owning
Build on a framework yourselfTeams with Python/.NET engineers and time to learn agent-specific failure modesEngineering months + the eval layer nobody budgets forEverything, including the maintenance
Buy an agent platformStandard workflows, no engineering capacity, need it live this weekPer-seat or per-run fees forever; their roadmap, not yoursA subscription
Delivery partner (us)You know the workflow, want it in production in weeks, and want to own the result$8K–$35K typical; 2–4 weeks to a working agentCode, prompts, orchestration, evals — handed over

We don't resell a framework. We have no partnership that pays us to recommend one, which is the only reason our recommendation is worth anything. When a platform tool is genuinely the better fit, that is what we will tell you.

What we can show you: we run our own agents in production on this site — a sales agent, an appointment-booking agent and an email agent. We would rather show you those working than hand you a case study.

The number behind the warning

What does it actually cost to change framework later?

Between two and eight engineering weeks for a working single-agent system, depending on how much of your logic leaked into framework-specific code. The framework itself is rarely the expensive part.

This is the number nobody puts in the proposal. So here it is, before you commit to anything.

What movesEffortWhy
Prompts and tool definitionsDaysMostly portable if you kept them out of the framework's classes
Orchestration & state model1–3 weeksThe real rewrite — graphs, crews and handoffs do not map onto each other
IntegrationsDays–1 weekCheap if they sit behind your own interfaces; painful if not
Evaluation harness1–2 weeksOften rebuilt from scratch, because it was tied to the old runtime
Re-earning trust internallyUnbudgetedThe one nobody puts in the estimate, and the one that hurts

How we keep that number small on builds we run: business logic lives in your code, not in the framework's abstractions. Tools sit behind plain interfaces. Evals run against the agent's behavior, not its internals. None of that is clever — it is just refusing to marry the framework.

HOW A FRAMEWORK CHOICE QUIETLY BECOMES A REWRITE

MONTH 1

The demo works. Three steps, happy path, everyone delighted

MONTH 3

Real users do things the demo never did

MONTH 5

It failed and the trace does not go deep enough to say why

MONTH 6

The rewrite. Not a bad framework — an unseeable one

The exit is at stage 2 — instrument it before the edges arrive, not after

Illustrative — the failure pattern behind the 88% figure, not a specific client’s project.

Our process

How do we pick the framework for your build?

Four steps, and the first has nothing to do with frameworks. We look at the shape of the work, then at what has to be provable, and only then do we choose.

Match the task shape

Linear pipeline, branching workflow, or a set of roles? That one question eliminates most of the table above before we discuss anything else.

Decide what must be provable

What will you have to show a customer, an auditor, or your own board? That sets the evaluation design — and it often overrides the framework preference.

Build, integrate, deploy

A working agent on your data in 2–4 weeks, wired into your real systems rather than a sandbox copy.

Hand it over

Code, prompts, orchestration and evals become yours. Keep us for iteration if it helps — the agent runs either way.

You are never buying a black box from us. Step four is the point of the whole thing.

See it on your workflow before you payWe’ll build a working agent on your real process and show it running.
Get my free demo →

THE SWAP TEST · FIVE PARTS MOVE, ONE DOESN’T

Every decision must be reconstructable later

Shortlist: graph-based control

Specialists hand work to each other

Shortlist: role-based orchestration

It must pass your security review

Shortlist: whatever your platform already runs

Retrieval is the product

Shortlist: document-pipeline engines

The reasoning model, the tools and the hosting stay swappable in all four branches

Illustrative — our selection logic, not a ranking. We do not resell a framework.

What it costs to build

How much does it cost to build an agent on one of these frameworks?

A focused production agent runs $8K–$35K to build, plus a monthly run cost. The framework is free; the engineering, integration and evaluation work is what you are paying for.

A representative $22,000 stateful agent (LangGraph)

Discovery & workflow mapping$3,000
14%
The real process, including the exceptions people forget to mention
Orchestration & state design$7,000
32%
The graph, durable state, retries, human-in-the-loop approval points
Evaluation & observability$5,500
25%
The layer 64% of failed pilots skipped
Integrations & tool wiring$4,500
20%
Your systems, behind interfaces we own rather than the framework's
Handover & documentation$2,000
9%
Code, runbook and training so your team can carry it
Representative total~$22,000

Representative scope for planning, not a past client invoice. Run cost lands around $400–$2,500/month depending on volume and model choice.

The cheapest quote you get will be the one that quietly dropped the evaluation line.

Notice which line is third-largest. Evaluation and observability is the piece cheap quotes leave out, and it is the same piece that 64% of failed pilots named as their blocker. When a quote comes in suspiciously low, that is usually the line that went missing.

The other half of the decision

Which LLM should the agent actually run on?

The framework is the wiring; the model is the engine, and it is where nearly all of your running cost goes. Every framework on this page is model-agnostic, so this is a separate decision — and a reversible one, if you build it that way.

The framework is a decision you make once. The model is a bill that turns up every single month.

ProviderWhere it is strongestWhere it costs you
Anthropic (Claude)Long multi-step tool use, following a policy exactly, honest “I don’t know” behaviorTop tier is the most expensive per token on this table
OpenAI (GPT)Broadest ecosystem and tooling; the cheapest small models by a wide marginThe research-grade tier is dramatically more expensive than everything else
Google (Gemini)Cheapest flagship-class option; natural fit if you are already on Google CloudFewer mature agent integrations than the other two
Open-weight (Llama, Mistral, Qwen)No per-token cost, full data control, runs on your own hardwareYou pay in GPUs and engineering time instead — rarely cheaper below serious volume

What it costs per million tokens

ModelInputOutputContext
Claude Opus 5$5$251M
Claude Sonnet 5 — our usual default$3$151M
Claude Haiku 4.5$1$5200K
GPT-5.5$5$30
GPT-5 mini$0.25$2
Gemini 3.1 Pro$2$12
Gemini 3 Flash$0.50$3

Claude figures from Anthropic’s published pricing; OpenAI and Google figures from 2026 pricing aggregators. List prices move — we re-check them at every build, and so should you before budgeting.

The mistake we would avoid: picking the most capable model for every step. Most agents do one hard thing and twenty easy ones. Route the easy steps to a cheap model and the hard step to a strong one, and the bill usually falls by more than half without touching quality. That routing lives in your code — which is exactly why we keep it out of the framework’s abstractions.
Not sure which model your workflow needs?We will benchmark two or three on your actual task and show you the quality and the cost side by side.
Compare models on my task →
Pricing

See the range, then see where you'd land

Same anchors as every LoopHawk agent build. No per-seat license, no lock-in.

No discovery fee buried in the footnotes, no per-seat surprise later. You see the range before you ever speak to us.

Proof of concept

Starter agent

from $1,800
  • One job, one framework
  • Working demo on your data
  • ~1–2 weeks
Most framework builds
Ready for real traffic

Production agent

$8K–$35K
  • Durable state & retries
  • Evaluation + observability
  • Integrations & handover
  • Live in 2–4 weeks
Multi-agent

Custom / multi-agent

from $40K
  • Coordinated agents, A2A
  • Deep integration work
  • Security review included

Ongoing support from $200/mo. Migrations off a maintenance-mode framework are quoted from the table in what a rewrite costs.

Who has to run it at 2am

Do the right AI agent frameworks change with the size of your team?

More than your industry does. The AI agent frameworks that survive a two-person team are not the ones that survive a platform team. What breaks a two-person team and what breaks a fifty-person team are different failures, and the frameworks are not equally good at surviving either one.

Two engineers and no platform team

Pick the one with the least to operate. Every hour spent running orchestration is an hour not spent on the thing customers pay for. A managed runtime or a low-code layer is not a compromise at this size — it is the correct call.

Typical build: from $1,800

A product team with real traffic

Now the control model matters, because you have enough concurrent runs to meet the edge cases a demo never reaches. This is where graph-based control earns its complexity, and where you can no longer skip the tracing layer.

Typical build: $8K–$35K

Platform team, security review, compliance calendar

Choice narrows to whatever your security team will sign off, whatever your existing telemetry can already read, and whatever survives an audit. That is a far shorter list than the internet suggests.

Typical build: from $40K · ongoing from $200/mo

Match the framework to whoever has to run it at 2am.

The honest answer

Does your industry change the answer?

Less than vendors imply. Industry rarely changes which of the AI agent frameworks wins. It changes which of the six parts you are not allowed to get wrong.

IndustryWhat changesWhat it rules out
E-commerceLive order state — the agent must read the truth, not a cached copyAnything that cannot call your order system mid-conversation
Healthcare & financeEvery decision must be reconstructable months laterLoops you cannot replay step by step
LogisticsLong-running jobs that outlive the requestFrameworks with no durable checkpoint
Professional servicesDocument pipelines carry the valueThin retrieval layers
SaaSMulti-tenant isolation and per-customer configAnything with global mutable state

Your industry does not pick the framework. It picks the part you cannot compromise on.

Straight answer

Should you build it yourself or have us build it?

If you have Python or .NET engineers with time to learn agent-specific failure modes, build it yourself — the frameworks are open and good. Hire out when the deadline is real and the learning curve is not the thing you are being paid to climb.

We turn work down on this basis. If your team can do this in two weeks, we would rather say so than sell you a project.

✓ Have us build it if

  • you need it in production in weeks, not quarters
  • nobody in-house has shipped an agent before
  • a wrong answer has a real cost — money, compliance, trust
  • you're stuck on a maintenance-mode framework
  • the prototype works and production doesn't

✗ Go DIY if

  • your team already runs LLM systems in production
  • the scope is small and the blast radius is tiny
  • learning the stack is the goal
  • you have more engineering time than budget
When the question changes

What should you look for in whoever builds this for you?

At some point the question stops being which of the AI agent frameworks and starts being who is going to run this in production. Six things worth checking — and the last one is the only one that cannot be faked in a sales call.

  • They name a framework before they scope your problemThat is a reseller, not an advisor. The recommendation should arrive after the requirements, never before them.
  • They can explain the control model in your wordsIf the graph-versus-role difference only survives in jargon, they have not built enough of them.
  • They price the build, not the licenseThe framework is free. What costs money is proving it works, wiring it in, and the on-call rota afterwards.
  • They tell you when not to buildThe honest answer to “do we need multi-agent” is usually no. A partner who never says no is optimizing for scope.
  • They hand over something you can operateIncluding the tracing layer, the evaluation set and the runbook — not just a repository.
  • They show you a working one before you payEverything above is a claim until something runs on your data.

Five of these are checkable in one call. The sixth is why we lead with a demo.

If you only have ten minutes

How do you choose a framework without a three-week evaluation?

Answer four questions in order. The first one that gives you a clear answer usually settles it, and you can stop.

Ten minutes on the right four questions beats three weeks of proof-of-concepts. We have watched both play out.

1. Does the work need to survive a restart?

If a run can take hours, needs approval partway, or must be replayable for an auditor, you want durable state. That points at LangGraph, or Microsoft Agent Framework if you are a .NET shop. If a failed run just means running it again, skip to question two.

2. What language does your team actually ship in?

Python teams have the widest choice. .NET teams should look hard at Microsoft Agent Framework, because the alternative is running Python services your team will not enjoy owning. TypeScript product teams are usually happier with Mastra or the Vercel AI SDK than bolting on a Python backend.

3. Are you already inside one vendor's ecosystem?

Deep in Azure, AWS or Google Cloud? The native option removes a category of integration work — Microsoft Agent Framework, Strands, or Google ADK respectively. The convenience is genuine. So is the trap underneath it. Trade it knowingly.

4. Are you still proving the idea?

Then optimize for speed to first result, not for architecture. CrewAI or a minimal loop like Agno will tell you in days whether the idea holds. Rebuild properly once it does — that rebuild is cheaper than the six months you would spend designing for a product nobody wanted.

What we would not do: pick based on GitHub stars, or on a benchmark run by whoever makes the framework. Neither predicts how the thing behaves at 2 a.m. in month four.
Before you commit

What should you check before you build on any framework?

Seven checks, each one learned the expensive way. Run them before the first sprint, not after the demo goes well.

Print it. Take it to every vendor you are talking to — us included — and make them answer each line.

CheckWhat good looks likeThe risk if you skip it
Release cadenceMeaningful commits and releases in the last 90 daysYou inherit a project in maintenance mode without being told
State & durabilityRuns survive a process restart without losing contextEvery crash becomes a support ticket and a manual replay
Human-in-the-loopA first-class way to pause for approval and resumeYou bolt it on later, badly, under compliance pressure
Observability hooksTraces you can read without a vendor dashboardThe 64% blocker — you cannot prove the agent is right
MCP supportNative or a maintained adapterCustom glue for every internal tool you connect
Escape hatchYour logic sits in your code, not in their classesA framework change becomes a rewrite (see the cost table above)
License & hostingPermissive license, self-hostable, no phone-homeA procurement surprise two weeks before launch

Run those seven against any framework on the map and the shortlist collapses to one or two fast. It is the same checklist we run on our own builds — there is no secret method behind it.

Want us to run these checks with you?Bring your shortlist. We'll go through all seven on a call and tell you what we'd pick.
Run the checklist with me →
Summarize this page with your AI assistant Open it in one tap — what changed in April, the current framework map, and what a build costs.
Questions we actually get

AI agent frameworks — your questions

What is an AI agent framework?

The code layer that controls how an agent plans, keeps state, calls tools and hands off to a human. Most 2026 shortlists come down to four names: LangGraph, CrewAI, Microsoft Agent Framework and the OpenAI Agents SDK.

Is there a single best AI agent framework?

No single one wins. LangGraph suits stateful production work, CrewAI suits fast role-based prototypes, Microsoft Agent Framework suits .NET and Azure estates, and the OpenAI Agents SDK suits teams already inside that ecosystem.

Is AutoGen still worth using?

Not for new builds. AutoGen moved into maintenance mode on 3 April 2026 when Microsoft Agent Framework 1.0 shipped. What you already run keeps running. Anything new should begin on the successor.

What happened to Semantic Kernel?

Same story as AutoGen. Its capabilities were consolidated into Microsoft Agent Framework 1.0 and the original moved to maintenance mode. It was not deleted, and it was not technically merged.

LangGraph or CrewAI — which should I start with?

CrewAI if you want something running this week and the run is short. LangGraph if the process has state worth protecting, needs human approval, or has to be replayable later.

What is an LLM agent framework, and is it different?

Same thing. "LLM agent framework" and "AI agent framework" describe the orchestration layer around a language model that lets it use tools and take multi-step action.

Do I need MCP or A2A support?

MCP matters if your agent calls internal tools and data — over 9,400 public servers now speak it. A2A matters when agents you do not own need to hand work to yours.

Do I need a multi-agent system?

Usually not at first. Around 22% of production deployments run three or more agents. One well-built agent with good tools is cheaper to run and much easier to debug.

What does a framework-based agent build actually cost?

Typically $8K–$35K for a production build, with a representative stateful agent around $22,000 and roughly $400–$2,500 a month to run. The frameworks cost nothing; the engineering around them is the spend.

How fast can a framework agent reach production?

Two to four weeks for a focused production agent, including integrations and the evaluation layer. Multi-agent systems take longer, mostly because of permissions and testing.

What does switching framework later cost?

Two to eight engineering weeks for a single-agent system. Orchestration and the evaluation harness are the expensive parts; prompts and tools usually port easily if you kept them out of framework classes.

Do we own what you build?

Yes. Code, prompts, orchestration and evaluation harness hand over to you. We build on open frameworks and resell nothing, so there is no license to lose.

How this actually goes

What happens after you get in touch?

Four steps. You see it working at step three, before any commitment.

HOW WE PICK YOURS — AND PROVE IT BEFORE YOU PAY

MAP THE WORKFLOW

Your real process, including the parts nobody documented

NAME THE LOCKED PART

The one thing that must not change in month six

BUILD A LIVE DEMO

On your data, free, before any commitment

HAND IT OVER

Repo, eval set, tracing and runbook — you own it

You own the build, the evals and the runbook · no license, no lock-in

Illustrative — our delivery sequence, not a specific client’s engagement.

You will have seen it run on your own data before you decide anything.

Get started

Skip the framework gamble — see it built on your workflow

Tell us what the agent has to do. We’ll name the framework we’d use and why, then build a working version on your real process so you can judge it before paying. If your team should build this in-house, we’ll tell you that instead.

You own the agent we build — open frameworks, no lock-in. We reply within one business day, and we’ll say so honestly if your team should build this themselves.
🇺🇸 US-registered · Model-agnostic · You own the code · Proven before you pay
Book a Free Call