GitHub Agentic Workflows changes model selection
Version 1.2.3 enables cost-first routing for eligible workflows, so AI configuration deserves a closer look.
The model you do not declare still matters
GitHub Agentic Workflows 1.2.3 introduces a change that is easy to miss: an eligible workflow with no declared model can now use model routing. Its default objective is cost, rather than preserving the exact selection behavior from before. If an automation relies on highly consistent results, this deserves a review before workflows are recompiled. (github.github.com)
That does not mean every workflow changes automatically. Default routing only applies when a workflow meets the router’s compatibility requirements and is not already setting a model, engine option, or another inference customization. When eligibility is not met, compilation keeps the previous behavior. (github.github.com)
Where this fits in a Laravel repository
A Laravel codebase usually has recurring work that does not need the same reasoning level as a code review that may open a pull request. Issue triage, label suggestions, CI failure summaries, and stale documentation checks are sensible candidates for an economical route. Work involving authorization, payments, migrations, or business rules deserves a more cautious route and a human review afterwards. (github.github.com)
The useful part is making that distinction visible in configuration instead of assuming one model is equally suitable for every job. GitHub Agentic Workflows supports model: router when the level should be selected per run, as well as router/economy, router/balanced, and router/robust when the workflow has a known risk profile. Those levels are selection targets, not guarantees that a run will succeed. (github.github.com)
A configuration for routine work
For a workflow that reacts to a new issue with an initial assessment, I would make the choice explicit. That avoids relying on an implicit default that could be adjusted again in a later release. The prompt still needs clear boundaries: inspect the available context, avoid fabricated causes, and treat the result as a suggestion rather than a verified fix.
---
on:
issues:
types: [opened]
permissions:
contents: read
issues: read
copilot-requests: write
engine: copilot
model: router/economy
safe-outputs:
add-comment:
---
Review the new issue, identify Laravel files that may be related,
and leave a comment with practical steps to reproduce the problem.
Do not suggest code changes without enough evidence in the repository.router/economy is intended for routine, low-risk jobs, while router/balanced is positioned for general-purpose work. For jobs where failure is more expensive, such as opening a pull request, the documentation describes router/robust as the highest target level. None of this replaces tests, protected branches, or code review. (github.github.com)
Keeping the previous behavior
When I do not want an eligible workflow to enter automatic routing, the documented setting is engine.model-routing: false. It is useful while comparing results, debugging an automation, or retaining the earlier model-selection behavior without pinning a specific model name. Declaring a fixed model remains an option when a task needs a particular capability or more predictable behavior. (github.github.com)
---
engine:
id: copilot
model-routing: false
---I would not add that setting everywhere by default. First, I would separate mechanical workflows from workflows that create changes and from workflows that touch sensitive information or product decisions. Routing is most useful when the chosen level matches the actual risk of the automation.
The small print before enabling it
Model routing is experimental, and GitHub states that its behavior and configuration may change. The same task may be assigned a different model or reasoning effort on another run, so it is not a good fit for strict reproducibility unless other variables are tightly controlled. If the router cannot select a candidate, the workflow fails rather than automatically reverting to a fixed model. (github.github.com)
It is not a complete spending control either. The router does not account for a run’s AI Credits budget when making its selection, and robust runs can consume more credits. If cost matters, the selected level should be paired with execution limits and the auditing tools provided by the platform. (github.github.com)
Another detail is that routing chooses the main agent’s model, not every sub-agent or internal component. In Copilot CLI mode, a sub-agent from a different model family can fail when it does not match the family selected for the main agent. Before applying routing to a complex workflow, I would inspect sub-agents, model policies, and audit output from a real run. (github.github.com)
A low-risk way to adopt it
I would start with a maintenance workflow that only reads the repository and posts an observation: issue triage, label recommendations, or documentation checks. router/economy makes sense when the scope is narrow, while router/balanced is a safer default when the agent has to connect several parts of the project. I would reserve router/robust for cases where a poor decision clearly costs more than the run itself. (github.github.com)
Then I would check which model was actually selected and what the run cost through gh aw audit <run-id> and gh aw logs. The goal is not to choose the cheapest model every time; it is to avoid paying for complex reasoning when the task only classifies, summarizes, or flags something a person will validate later. That separation becomes more valuable as a repository collects more maintenance automations. (github.github.com)
Source: Weekly Update – November 16, 2026 Official documentation: Model Routing