Replaying an Agent Framework run with the model server switched off

[xizhuomengcontin]

Verified against the current Python agent-framework package today: a ChatAgent run can be recorded from outside the process and replayed with no model server present.

Disclosure: I work on the recording tool (OrcaReplay, Apache-2.0, no paid tier).

The run

c = OpenAIChatClient(model="qwen3.5:0.8b")     # origin from OPENAI_BASE_URL, nothing in code
a = c.as_agent(instructions="Be terse.")
r = await a.run("Reply with exactly: MSAF")

$ orca record generic-openai -- python agent.py
agent-framework said: MSAF
info recorded run=run_319d42f446bb events=6 blobs=1 exit=0

$ taskkill /F /IM ollama.exe          # model server gone
$ orca replay last
agent-framework said: MSAF
info replay.done reused=1/1 exact=1 divergences=0 unmatched=0 exit=0

OpenAIChatClient picks the origin up from OPENAI_BASE_URL, so nothing is installed into the framework and no code changes.

One API note

Two things I got wrong first, in case they are worth a docs line — the constructor takes model=, not model_id=, and the agent factory is as_agent(), not create_agent(). Several examples floating around use the other spellings.

What it is for

Re-running an agent while you change the instructions, the tools or the middleware: model responses stay fixed, so the difference you see is your change rather than the sampler. Also CI without keys or spend, and chasing one bad run repeatedly for free. From a checkpoint the run can instead continue live on a different model.

Scope

egress=blocked on replay means model-provider egress. Recorded tool calls still execute for real, so it is not a sandbox. And it captures the HTTP to the model, not the framework's internals — not a substitute for its OpenTelemetry support.