Picking a coding agent feels like picking an IDE. It isn't. An IDE never held your team's working context, and switching editors never meant rebuilding how you deliver. You avoid coding agent lock-in by keeping instructions, knowledge, and integrations in your repo and in open standards, by confirming you can export what the agent produced, and by pricing the exit before you commit.
Lock-in builds up in the agent harness. The harness is the software wrapped around the model: it reads files, calls tools, keeps context, and asks for approval. (We cover what the harness has to do for long-running agents separately.) Here are the five portability checks we run before a team commits its delivery workflow to one.
1. Where do your agent instructions live? If your agent's behavior sits in a config file only one tool reads, you'll rewrite it when you leave. Keep the real instructions in a format many tools read, such as AGENTS.md, an open format for guiding coding agents that the Agentic AI Foundation now stewards under the Linux Foundation. Make each tool-specific file a thin pointer to it. The knowledge should live in the repo you own.
2. What survives a switch? Conversation history, learned patterns, and behavioral tuning don't travel. Assume you lose all of it. The defense is boring: put durable knowledge in the repo as docs, tests, and architecture notes the next agent can read cold. If your codebase only makes sense to the agent that's been living in it, you're already locked in.
3. Are your integrations built on a standard or a vendor feature? A tool connection built as a Model Context Protocol (MCP) server works in any harness that supports MCP. The same connection built on one vendor's own feature works in exactly one. On August 19, 2026, Cursor shipped subscriptions, pinned custom modes, subagents on their own virtual machines, and a /goal command in a single release. Every one of them is useful. Every one is surface you can't take with you. Count how many your workflow would depend on.
4. Can you export what the agent work produced? That means evals, review rules, prompts that took weeks to tune, and run logs you'll want for audits. Confirm bulk export exists before you sign. A vendor who can't answer this in writing has answered it.
5. What does leaving cost, priced today? Run the exercise before you commit. If you switched harnesses in six months, what would you rebuild, and how many engineer-weeks is that? If the number scares you now, it'll be worse after a year of buildup. Sometimes the proprietary features are still worth it. Fine. Make that a decision, not a discovery.
The pattern behind all five: vendors are competing to own the loop, and every convenience that deepens the loop raises the exit cost. You don't have to refuse the convenience. You have to price it.
We make this call on live delivery work in our AI implementation practice. The teams that stay portable are the ones that decided to before the first commit.