Picking an AI agent development platform is the wrong first decision.

That's an awkward way to open a list of platforms, so here's the evidence. In October 2025, OpenAI launched AgentKit, a visual canvas for building multi-agent workflows. On 3 June 2026 it deprecated the thing, with a shutdown date of 30 November and a recommendation to go back to writing code with the Agents SDK. The same deprecation notice retired OpenAI's own Evals platform and pointed users at Promptfoo, a third-party open-source tool. The Assistants API shut down for good on 26 August 2026, with the Responses API as the migration path.

OpenAI isn't an outlier. Microsoft killed both AutoGen and Semantic Kernel to ship a single Agent Framework 1.0 on 3 April 2026. Google retired the Vertex AI brand entirely on 22 April 2026, folding it into the Gemini Enterprise Agent Platform, and stated that all Vertex AI services and roadmap evolutions would be delivered exclusively through the new platform.

So any list that ranks these tools one through ten is out of date before it's indexed. What holds still is the shape of the decision: which layer of the stack you're actually buying, and what it costs you when the product on top of it gets renamed. That's how this list is organized.

What is an AI agent development platform?

An AI agent development platform is software that runs, orchestrates or governs autonomous agents: systems where a language model plans a sequence of actions, calls tools to execute them, and adapts based on results. The term covers four distinct product categories that vendors market interchangeably: agent harnesses, managed runtimes, orchestration frameworks, and control planes.

The conflation matters commercially. Gartner coined the term agent washing in June 2025 for "the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities," and estimated that only around 130 of the thousands of vendors claiming agentic capability were real.

Menlo Ventures put a number on the buyer side of the same problem. In its December 2025 survey of 495 US enterprise AI decision-makers, only 16% of enterprise deployments and 27% of startup deployments qualified as true agents. The remainder were routing logic or fixed sequences wrapped around a single model call.

The distinction isn't academic once an agent has a surface a person interacts with. We've written separately about designing interfaces for agentic-as-a-service products and robots, and the platform you pick constrains what that interface can honestly promise.

The four layers you're actually choosing between

Every tool in this article sits in one of four layers, and the layer determines how expensive it is to change your mind later. Evaluation and observability cut across all four rather than forming a layer of their own, which is why so many teams end up with none.

The harness is the layer most teams don't know they're choosing. Philipp Schmid's January 2026 essay gives the cleanest definition available: the harness is "the infrastructure that wraps around an AI model to manage long-running tasks": effectively an operating system between the model and your application. All four hyperscalers plus LangChain now ship one under that name, and none of them did a year ago.

Best AI agent orchestration frameworks in 2026

Orchestration frameworks are the layer to start with if you're running your own infrastructure. They shape control flow (how agents hand off, how state persists, where a human approves), while leaving execution to you. All the options below are open source, which makes the choice reversible in a way the runtime layer isn't.

The first table covers what each framework is. The second covers where each one fits.

Two things stand out reading those two tables as a buyer rather than an engineer. The first is that OpenAI's agent library, the one it told Agent Builder customers to migrate to, is the only major framework here that hasn't shipped a 1.0. The second is that Microsoft's is the only one with a genuinely permissive license and a hyperscaler behind it, which is worth something if you're writing a five-year procurement case.

Best managed agent runtimes in 2026

Managed runtimes take over process execution and sandboxing. You stop operating long-running agent processes and start paying someone per vCPU-hour to do it. The trade is real: this is the layer where the vendor's session semantics become your architecture.

Google hasn't published GA-versus-preview status for the individual components of its platform, and the docs expose both v1 and v1beta1 API versions, so treat any specific component as preview until Google says otherwise. Anthropic hasn't published Managed Agents pricing at all.

Temporal deserves a note because it's the odd one out. It isn't an agent platform, it's durable execution, and it raised $300M at a $5B valuation in February 2026 largely on agent workloads, naming OpenAI, Replit, Lovable, ADP and Block among its customers. When the companies building agent platforms run their own agents on something else, that's information.

What's an agent harness, and why does it matter more than the framework?

An agent harness is the code that wraps a model to run long tasks: it manages the loop, compacts or offloads context, decides which tool calls need approval, spawns subagents, and holds session state. It's the layer that determines whether your agent survives a four-hour job, and it's the hardest layer to swap out because your prompts and tool definitions are written to its assumptions.

The current options, all shipped within the last eighteen months:

• Claude Agent SDK (Anthropic). Claude Code's own loop as a library: tools, hooks, subagents, permissions, sessions, skills. Python and TypeScript only; other languages shell out to the CLI. Governed by Anthropic's commercial terms rather than an open-source license.

• Harness Agent (Microsoft). Planning, todo tracking, context compaction and tool approval, inside Agent Framework 1.0.

• AgentCore harness (AWS). Two API calls, CreateHarness and InvokeHarness, GA since 18 June 2026.

• Deep Agents (LangChain). An open-source harness for long-running tasks with rubric-based self-evaluation. Still alpha as of March 2026.

Anthropic's own April 2026 engineering write-up on scaling Managed Agents is the most useful public document on what breaks at this layer, because it's retrospective rather than promotional. Container failure meant total session loss in coupled architectures. Sessions got stuck with no debugging visibility. And they named a failure mode you won't find in any vendor deck: context anxiety, where a model senses the context limit approaching and prematurely wraps up a task it hasn't finished. Their fix, a durable session event log independent of the context window plus stateless harnesses that recover via wake(sessionId), cut p50 time-to-first-token by roughly 60% and p95 by more than 90%.

The line from that post worth pinning above a whiteboard: harness assumptions rot as models improve. The harness you wrote for last quarter's model is this quarter's liability. Choosing well at this layer is a large part of what AI Engineering actually means in practice.

Agent governance and control planes: the category that didn't exist last year

Agent control planes handle identity, registration, policy enforcement, audit trails and discovery of unsanctioned agents. This category went from nonexistent to generally available inside twelve months, which tells you where enterprise agent projects were actually getting stuck.

Microsoft Agent 365 reached general availability on 1 May 2026 at $15 per user per month, or bundled in Microsoft 365 E7. It ships an agent registry, agent identity, network controls, runtime blocking, cross-platform registry sync with AWS Bedrock and Google Cloud, and shadow-AI discovery that specifically detects locally-running agents including OpenClaw, GitHub Copilot CLI and Claude Code. Google's equivalent components (Agent Identity, Agent Registry, Agent Gateway and a security dashboard) arrived with the Agent Platform announcement in April 2026.

When a hyperscaler builds detection for a class of software, that class has become a governance problem. Deloitte's April 2026 survey of 3,235 IT and business leaders across 24 countries found that only 21% have a mature governance model in place for agentic AI, and named exactly what's missing: clear boundaries defining which decisions agents make independently versus which require human approval, real-time monitoring that flags anomalies, and audit trails.

Best AI agent evaluation and observability tools, and who owns them now

Evaluation tooling is the layer most agent projects skip and the one that decides whether they ship. It's also the layer consolidating fastest, which makes vendor longevity a live selection criterion rather than a paranoid one.

Three acquisitions in sixteen months, all by incumbent observability or networking vendors. Two of the three announcements don't name the specific product you'd be buying. If you're standardizing on an eval vendor this quarter, ask whose roadmap it sits on in 2028.

Are no-code AI agent builders worth it?

No-code and low-code agent builders are the fastest way to a working agent and the fastest way to a dead product line. The two options worth naming are n8n, which raised $180M at a $2.5B valuation in October 2025, and Zapier Agents, which reaches across 9,000-plus apps and meters usage in "activities."

Two caveats, both concrete. n8n is not open source, despite being widely described that way. It ships under a Sustainable Use License restricting use to internal business purposes, with redistribution allowed only free-of-charge and non-commercially, and .ee. files requiring an enterprise license. And Zapier doesn't publish per-unit pricing for agent activities on its product page, so model your costs against a real workload before committing.

The structural caveat is the one from the top of this article. Visual builders shipped by model labs have a demonstrably shorter half-life than the SDKs underneath them: OpenAI's lasted eight months. Independent workflow vendors like n8n have survived considerably longer, because the product is the business rather than a funnel into a model API.

What actually breaks when agents reach production?

Agent projects rarely fail on framework choice. They fail on evaluation, cost, context and permissions, and there's now enough published research to name each failure precisely rather than gesturing at "reliability."

Evaluation coverage doesn't match the risk being taken. Amplify Partners' 2026 AI Engineering Report, run with Notion and Vercel across 1,000-plus AI engineers, found that among teams deploying agents the share granting write permissions rose from 52% to 89% in a year. LangChain's State of Agent Engineering, surveying 1,340 professionals, found 89% have implemented observability but only 52.4% run offline evaluations and 37.3% online. Near-universal write access against roughly half-coverage on evals is the headline tension of the whole category. It's the same trade-off we've written about in the narrower case of the benefits and risks of vibe coding, scaled up to systems that act on their own.

Context degrades before the window fills. Chroma's Context Rot report tested 18 models across Anthropic, OpenAI, Google and Alibaba, isolating input length while holding task complexity constant. The finding: models "do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows." Distractors amplify the decline, and low semantic similarity between question and answer accelerates it. Anthropic's context engineering guidance is the practitioner counterpart, prescribing compaction, structured note-taking to external files that survive context resets, and subagents returning condensed summaries.

The token bill is not linear. Anthropic's multi-agent research post puts real multipliers on it: agents use roughly 4× the tokens of a chat interaction, and multi-agent systems around 15×. The honest part of that post is the next clause: token usage alone explains 80% of the variance in research task performance. You're buying quality with tokens, and the ROI model has to say so.

Security stopped being hypothetical. OWASP published its Top 10 for Agentic Applications on 9 December 2025, grounded in 2025 incidents rather than projections. The Cloud Security Alliance's April 2026 research note on indirect prompt injection reports that Google documented a 32% relative increase in malicious injection content between November 2025 and February 2026 across billions of crawled pages, with Palo Alto's Unit 42 documenting 12 real-world cases across 22 delivery techniques. A follow-up note in July 2026 puts it plainly: the Model Context Protocol tool description field "is an unsanitized attack surface." The MCPTox benchmark measured a 36.5% average attack success rate across 20 models.

And capability plateaus short of the business task. Salesforce AI Research's CRMArena-Pro benchmark ran 19 expert-validated tasks across sales, service and configure-price-quote workflows. Leading agents scored around 58% single-turn, collapsing to roughly 35% multi-turn. The finding that should concern anyone deploying into a customer-facing process: all models showed near-zero inherent confidentiality awareness, and prompting to fix it degraded task completion.

Every one of those failures lands in the same place: the interface between what the agent decides and what a person approves. That's a design problem as much as an engineering one, which is why we built an agent orchestration UX framework rather than treating supervision as a settings screen.

Gartner's position reads consistently with all of the above. Its April 2026 Hype Cycle for Agentic AI places agentic AI at the Peak of Inflated Expectations, with 17% of organizations having deployed agents and more than 60% expecting to within two years (the most aggressive adoption curve among the technologies Gartner measures), while noting that "most deployments remain narrowly scoped, and fully autonomous agents are not ready for the majority of enterprise use cases." Its widely-quoted prediction is that over 40% of agentic AI projects will be canceled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." Worth reading that as written: canceled, not failed, and a prediction rather than a measurement.

How we run agentic development, and where the gate sits

We build with these tools daily, and the thing that made the difference for us wasn't the platform. It was deciding where a human has to say yes.

At an internal engineering craft session in July 2026, one of our Principal Engineers, Rui Saraiva, walked the team through the Agentic Workflows approach we use on client product work. The shape of it is simple: the agent plans before it builds, a person reviews and challenges that plan, and nothing merges without a second human gate at pull request review. Forty-eight minutes in, the questions were still coming.

The principle we landed on is one sentence: AI proposes, people decide. Which sounds like a slogan until you read Amplify's 89%-write-permission number next to LangChain's 52%-eval-coverage number and realize most teams have inverted it.

That's the layer we work in as an AI-native Digital Product Studio: not building agent platforms, but choosing among them and wiring the review gates that make their output trustworthy. It's the practical end of why we're becoming an AI-native Digital Product Studio, and it's also the least glamorous part of the job. Nobody's writing listicles about approval workflows.

How to choose an AI agent development platform

Choose by layer and by constraint, not by feature grid. The four questions below narrow the field faster than any comparison table, including the ones above.

Two selection criteria that don't appear on vendor comparison pages and probably should. First, has this product shipped a 1.0? OpenAI's Agents SDK, the recommended migration path off a deprecated product, is still on 0.Y.Z versioning where minor releases can break you. Second, who will own this vendor in 2028? That question was rhetorical two years ago. After Weights & Biases, Galileo and Arize, it isn't.

Both questions are really about the same thing: how much technical debt you're agreeing to take on at each layer, and whether you'd be able to service it when the vendor changes direction. If you can't answer that for your own stack yet, an AI Readiness Assessment is the cheaper first step than a platform decision.

The protocol layer is the one piece of this stack that has genuinely settled, and it's worth building against. Anthropic donated the Model Context Protocol to the new Agentic AI Foundation under the Linux Foundation on 9 December 2025, co-founded with Block and OpenAI, with AWS, Google and Microsoft among the platinum members. Google's Agent2Agent protocol went to the Linux Foundation in June 2025 and reached v1.0, past 150 participating organizations. Tools churn. MCP and A2A look like the parts you can bet a roadmap on.

Why the review gate outlives the platform

The uncomfortable read on this list is that most of it will be wrong within a year. Two of the four categories didn't exist in mid-2025. Two more named deprecations land on 30 November 2026. Three eval vendors changed hands while we were writing.

What hasn't moved is the thing the research keeps pointing at from different directions. Chroma's context rot, Anthropic's context anxiety, Salesforce's confidentiality gap, Deloitte's 21% governance maturity, Amplify's 89% write permissions: these are all the same finding wearing different clothes. Agents fail at the boundary where they're trusted to act without a human checking the reasoning. Every platform in this article is a bet on how you manage that boundary.

So evaluate the tools, and pick the layer deliberately. Then spend the rest of your effort on the part that no vendor sells you: knowing exactly which decisions your agents get to make alone.