plugin

Evolving Agents — a plugin for the Claude Agent SDK

The SDK already runs the loop, ships the tools, manages context, and enforces permissions. This plugin does not re-implement any of that. It adds the part the SDK leaves open: what happens to an agent's history once it has more than one.

What the SDK gives you, and where it stops

fork branches a session. Nothing merges the branches back. There is no diff between two sessions, nothing refuses to promote an agent whose eval fails, and the transcript is plain JSONL that anyone can hand-edit. Sessions also persist the conversation, not the agent — the skills, subagents, model and goal that produced it are not versioned alongside it.

	Agent SDK	This plugin
Branch a session	fork	—
Compare two sessions	—	avcs_diff
Rejoin two branches	—	avcs_merge, with a --reconcile seam for the non-code dimensions
Refuse to ship a regression	—	avcs_freeze, which fails unless the eval passes
Prove a transcript wasn't edited	—	Ed25519-signed commits
Keep the full trace past compaction	summarised away	archived by the PreCompact hook

Install

from claude_agent_sdk import query, ClaudeAgentOptions

options = ClaudeAgentOptions(
    plugins=[{"type": "local", "path": "/path/to/evolving-agents/plugin"}],
)

import { query } from "@anthropic-ai/claude-agent-sdk";

const messages = query({
  prompt: "…",
  options: { plugins: [{ type: "local", path: "/path/to/evolving-agents/plugin" }] },
});

Confirm it loaded by reading plugins on the init system message.

What it installs

An MCP server (agentvcs) exposing the history as tools the agent can call on itself mid-loop: avcs_log, avcs_diff, avcs_merge, avcs_freeze, avcs_rollback, avcs_replay, avcs_branch, avcs_recall, avcs_eval, and more. Zero runtime dependencies — standard library only, and there is a test that fails if that ever stops being true.

Three hooks, none of which can block your agent:

• PreCompact copies the transcript to .agentvcs/sdk/transcripts/ before compaction summarises it. This is the only point where fidelity is lost and cannot be recovered afterwards.
• SubagentStop records which subagents ran, which is the swarm dimension a session transcript flattens.
• SessionStart notes the session so later commits can be traced back to it.

Every hook exits 0 on any input. A full disk, a read-only checkout or a malformed event is recorded and skipped, never raised into your session.

Where the data goes

.agentvcs/sdk/ in your working directory — events.jsonl, swarm.jsonl, and transcripts/. Plain files, readable without this plugin installed.

Scope

This plugin targets the Agent SDK and nothing else. The organisation's work on decode-time constraint (token-trie) and on model-internals research (sleep-harness) needs logit and weight access that an API-backed SDK does not expose; those live in their own repositories and are not part of this.