Swapping Codex in as an AGNT fleet adapter in 30 minutes

AGNT's fleet v2 architecture (shipped in v0.2) decoupled agents from model backends. An agent declares a preferred adapter — claude_local, codex_local, gemma_local — and the runtime resolves that adapter at spawn time. Swapping adapters is a hot-patch, not a restart.

We documented this path through a live swap: a content-writer agent running on claude_local was switched to codex_local in a single API-PATCH call. Total time from start to first completed task on the new adapter: under 30 minutes, most of which was setting up the Codex environment keys.

The adapter interface is narrow by design. Four methods: `complete(prompt, max_tokens)`, `stream(prompt)`, `supports_tools()`, and `health_check()`. Every model wrapper implements these and nothing else. Codex maps cleanly — its completion API has equivalent parameters, and the streaming interface translates 1:1.

Where it gets sharp: tool-calling semantics diverge. Claude's tool_use blocks are structured; Codex's tool calls are function-shaped. The adapter layer normalizes both into AGNT's internal ToolInvocation shape, but if you're writing a custom adapter, expect ~40% of the code to be schema translation.

The full walkthrough lives at /guides/ecosystem/swap-codex-as-fleet-adapter. It covers the API-PATCH request, the cwd preservation trick, the rollback path, and a short checklist for validating the swap before you trust it in production.