On June 3, 2026, OpenAI posted a quiet deprecation notice. Agent Builder, the drag-and-drop agent tool launched at its developer conference in October 2025, would stop working on November 30. Eight months, launch to shutdown. Any company that had wired its customer workflows into that canvas now owns a migration project it never budgeted for.

That’s the market you’re walking into when you pick a home for AI agents. So it’s worth slowing down before you commit.

Quick definitions first. An agent is AI that does things rather than just answering. Where ChatGPT types a reply, an agent books the refund, updates the record, and quietly files the ticket afterwards. A platform is where those agents live and run in production, the way an employee needs an office, a badge, and an IT department, not just a brain. And before comparing offices, it’s fair to ask whether you need the employee at all. There are plenty of jobs where agents are the wrong tool.

If you do need them, there are five realistic places to run them in 2026. Each trades convenience against control in a different ratio.

AWS Bedrock AgentCore, the utility company

Amazon’s offering went generally available on October 13, 2025, and it behaves like a utility. You don’t buy servers. You plug in an agent built with almost any framework, running almost any AI model, and AWS meters what it consumes. Published rates are $0.0895 per virtual CPU-hour and $0.00945 per gigabyte of memory per hour, billed on actual use rather than reserved capacity. Since agents spend much of their time waiting on slow external systems, that metering matters.

On lock-in, AgentCore is a strange split. It doesn’t lock you to a model or a framework, which is rare. It does lock you to AWS plumbing, the identity systems and network configuration that take months to set up and would take months to unwind. On data control and compliance it’s the strongest managed option here: your agents can run inside your private network, and the audit trail satisfies the kind of reviewer who asks hard security questions of every vendor. It’s built for enterprises already living on AWS, and it feels like it. Small teams will find the setup heavy.

OpenAI, fast and prone to moving

OpenAI gives you the quickest path from idea to working agent. It also renovates while you’re living in the building. The Assistants API, once the flagship way to build agents on OpenAI, shuts down on August 26, 2026, according to the company’s own deprecation schedule. Its replacement is the Responses API, and the recommended way to build agents is now the Agents SDK, an open-source code framework. Agent Builder, as noted above, lasted eight months.

None of this means the technology is bad. The models are excellent and the developer experience is polished. But the pattern is clear: you’re building on ground that gets re-graded roughly yearly. Lock-in is high, because everything you build assumes OpenAI’s models and OpenAI’s data handling. Costs are per token, the small chunks of text AI models read and write, which is easy to start with and hard to predict at scale. This is the right choice for teams who value speed above all and accept that some rework is the rent.

Claude and MCP, the standard plug

Anthropic’s approach centers on MCP, the Model Context Protocol (an open standard that lets any AI model plug into any tool or data source through one common format, the way USB lets any device plug into any laptop). Anthropic released it in November 2024, and in December 2025 donated it to the Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI. Google adopted it too. That donation is the important part. A connector you build to your billing system today speaks a language every major AI vendor now understands.

So the lock-in math changes shape. You still depend on Anthropic’s models, and Claude’s per-token pricing (mid-tier Sonnet models list at $3 per million input tokens and $15 per million output) is comparable to rivals. But the expensive part of agent work, the integrations into your own systems, becomes luggage you can carry to another vendor. Data control is contractual rather than physical: your data flows through Anthropic’s API under its retention terms, or through AWS and Google Cloud versions of Claude if you need it inside your cloud. Who it’s for: teams that expect to still be running agents in five years and want the integration work to survive whichever model wins.

LangGraph and the open-source route

LangGraph, from the LangChain team, is an orchestration framework, which is a fancy term for the software that decides which step an agent takes next and what happens when a step fails. The library is open source and free. You run it on your own machines, point it at any model, and nothing about it phones home. Data control is total because the data never leaves you.

The costs arrive elsewhere. The company sells managed services around the free core: LangSmith, the observability product that records what your agents actually did, starts free at 5,000 traces a month and moves to $39 per seat per month on its Plus plan, and hosted LangGraph execution bills around $0.001 per node executed past the free allowance. The bigger cost is people. Open source hands you a professional kitchen instead of a restaurant meal, and someone on staff has to know how to cook. Compliance is whatever you build, which auditors treat as a question, not an answer. This route fits companies with genuine engineering depth who want leverage over every layer.

Fully custom, for when the agent is the product

The last option skips frameworks entirely: direct calls to a model’s API wrapped in your own loop. Here’s the industry’s open secret. The core of an agent is maybe a hundred lines of code. The model proposes an action, your code executes it, the result goes back to the model, repeat. Everything hard lives around that loop: permissions, retries, monitoring, cost caps, and the memory of what the agent already tried.

Custom is the maximum on every control dimension and the maximum on cost, because the cost is salaries. One senior engineer runs $200,000 or more a year in the US, which buys a lot of managed platform. The full budget math for building AI products in 2026 is its own topic, but the short version is that custom only makes sense when the agent is your actual product, when its behavior is your competitive edge, and when a platform’s limits are visibly costing you customers. For an internal helpdesk agent, it’s a vanity project.

The part that outlasts every platform

Here’s the honest ending, and it cuts against the premise of the comparison. The platform choice matters less than two assets that no platform sells you.

The first is your evals. An eval is a repeatable test that scores your agent’s output against examples you already know are right, like a driving test you can re-run every week. Fifty real cases from your business, with known correct answers, checked automatically. Teams with evals can switch platforms in weeks, because they can prove the migrated agent still works. Teams without them are guessing, so they stay put, and that fear is the deepest lock-in of all. Notice that OpenAI deprecated its own hosted Evals product alongside Agent Builder. Even your test infrastructure shouldn’t live on someone else’s roadmap.

The second is your data. The logs of what your agents did, which answers users accepted, which ones a human corrected. That record is what makes your agent better than a competitor’s copy of the same model, and it should land in storage you own, whatever platform produced it.

Pick the platform that fits your team today. Rent the office. But make sure the skills and the filing cabinet belong to you.