Latest in AI
GitHub Agentic Workflows Changed Models? Check Routing Before Your Prompt
The new no-model default can change how an agent chooses its model, while imports, runtime variables and compiled workflows can make two similar configurations behave differently.
GitHub Agentic Workflows now routes eligible workflows without an explicit model toward cheaper models on GitHub Copilot. If an upgraded workflow behaves differently, check its model configuration before rewriting the prompt. The same task instructions can now reach a different model-selection path.
GitHub’s November 16 update describes this change in v1.2.3, released November 13 and marked pre-release. Eligible configurations default to goal: cost and mode: auto. This is a migration check for teams adopting that behavior, not a claim that every existing scheduled workflow changed overnight.
Check the compiler and the deployed workflow separately
The versioning reference distinguishes the local or CI extension from the compiled .lock.yml file that Actions executes. The generated file embeds its compiler version. Start by checking both rather than assuming the version installed on your laptop explains a run created elsewhere.
gh aw version
Then inspect the workflow’s compiled file and its GH_AW_INFO_AWF_VERSION value. Keep the previous Markdown source, lock file and a representative run ID before preparing an upgrade branch. Recompile there and review the generated diff alongside the source change.
This is a useful boundary for incident diagnosis. If nobody changed the deployed artifact, a local tooling upgrade is not enough evidence to explain yesterday’s production behavior. Conversely, a small source diff can conceal a meaningful compiler-default change in newly generated output.
No model, auto and router are different choices
The model-routing reference still labels routing experimental. model: auto uses the engine or provider’s selection mechanism; model: router selects gh-aw’s per-task routing. They are not interchangeable spellings.
Default routing applies to eligible copilot, claude, codex and pi workflows using GitHub Copilot inference. Claude and Codex also need model-provider: github. Check imported workflows as well as the local file: a model, fallback configuration or model environment setting can prevent the no-model default from applying.
For Copilot, Claude and Codex, nonempty GH_AW_MODEL_AGENT_<ENGINE> or GH_AW_DEFAULT_MODEL_<ENGINE> repository or organization variables can override default routing at runtime. The agent-specific variable wins. Pi does not have those overrides.
That makes “these two files look identical” a weak diagnosis. Compare their imports and deployment configuration before treating different model choices as random behavior. Record the effective settings in the upgrade review so another engineer can reproduce the comparison without reconstructing your shell environment.
Opt out deliberately, without pretending to pin a model
For an existing Copilot engine block, GitHub documents this opt-out:
engine:
id: copilot
model-routing: false
Merge the setting into the existing block rather than replacing unrelated configuration. It disables this routing default without selecting a fixed model. If your requirement is a particular model, use an explicit supported model instead and review its availability separately.
I would separate two decisions here: accepting the tooling upgrade and accepting a different model-selection policy. A team can evaluate the new routing behavior in a test workflow without making every recurring maintenance task part of the experiment. Keeping those changes separate also makes a rollback easier to explain.
Compare useful outcomes, not just successful runs
GitHub’s audit command accepts multiple run IDs, using the first as the comparison baseline. Replace these placeholders with actual runs:
gh aw audit BASELINE_RUN_ID CANDIDATE_RUN_ID
Use comparable inputs rather than comparing a quiet Sunday report with a busy release-day report. For an issue-triage workflow, assemble a small review set containing a straightforward bug, a duplicate and an ambiguous report. Decide the expected labels and escalation behavior before looking at the candidate output.
A run that finishes without an error can still miss the duplicate or apply an unjustified label. Record those mistakes alongside latency and spending. If a cheaper run produces more cleanup work, its lower model bill is not yet a saving. This is a suggested evaluation workflow, not a benchmark I have run.
The routing reference also documents runtime restrictions. Explicit routing needs the AWF firewall and a supported sandbox; it is not supported with ARC DinD or a Docker-socket mount exposed to the agent. Check compatibility rather than weakening isolation to make an upgrade compile.
The useful question is whether the new policy preserves the decisions you trust this workflow to make. Until that is demonstrated, a smaller bill is an observation, not an upgrade verdict.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn engineering changes and their tradeoffs into clear stories developers can use.