The team behind the Agent Client Protocol (ACP), the open specification that lets code editors and coding agents speak a common language, published a draft of version 2 on July 20, 2026. The release rewrites four core parts of the protocol at once: how updates are sequenced, how messages and tool calls stream, how file changes get reported, and how permission prompts are structured.
The most consequential change removes the assumption that a coding agent works inside a single, user-initiated turn. In v1, a prompt request and its response implicitly bounded a unit of work, which confused developers building agents that keep working, queue tasks, or run background jobs after a user stops typing. v2 lets update notifications arrive whenever the agent has something to report, not only inside a turn, and a prompt response now simply confirms the agent received the message rather than marking the end of its work. The agent signals separately when it goes idle and is ready for new input.
Message and tool call updates now share one patching model across the protocol. An omitted field keeps its prior value, a null value clears it, a new value replaces it, and chunked content appends. Message IDs became mandatory, which lets an editor stream a response and later replace or redact it, a capability v1 offered only for tool calls.
File diffs got a full rework. The old two-field text diff is gone, replaced by structured operations that distinguish adding, deleting, modifying, moving, and copying files, with explicit handling for binary content. An agent can also attach a raw patch so an editor renders a diff directly instead of reconstructing one from the structured operations.
Permission requests, the prompts an editor shows before letting an agent run a command or touch a file, now carry their own independent label and explanation rather than inheriting one from the underlying tool call. That separation lets an editor describe a risky terminal command in plain language without also rewriting the tool call’s own display name.
ACP sits one layer away from MCP, Anthropic’s protocol for tool calling, which standardizes how a model reaches external tools and data. ACP standardizes the layer above that: how the editor itself (Zed, Neovim, a JetBrains IDE) hands off a session to an agent and gets structured updates back. As nearly every major vendor now ships its own coding agent, from GitHub to Cursor to internal enterprise tools, an editor that speaks one shared protocol avoids rebuilding custom integration plumbing for each new entrant.
The schema also builds in room to grow. Enum-style values across v2 accept unknown variants prefixed with an underscore, so a client running an older SDK will not break when it meets a newer agent’s custom extension. The project says the same pattern already proved out through more than a dozen proposals landed against v1 over the past year, using a review process it calls RFDs.
None of this is settled. The project describes v2 as a draft meant to surface implementer feedback before stabilization, and it is telling teams to gate v2 support behind feature flags and version negotiation instead of turning it on for everyone in production. v1 compatibility stays necessary too, since v1-only peers will not disappear the moment v2 ships.
Builders maintaining an agent, a client, or an SDK should treat the v2 schema as a target for prototypes and test suites this quarter, not for production defaults, and should plan to run both protocol versions side by side well into next year.
Published by the Agent Client Protocol project on July 21, 2026.
