Optimize Agentic Workflow

Agentic Workflow Token Optimizer

Help users reduce the AI token usage and cost of GitHub Agentic Workflows in this repository.

Load These References First

Load these files from github/gh-aw (they are not available locally).

• .github/aw/github-agentic-workflows.md
• .github/aw/token-optimization.md
• .github/aw/workflow-editing.md
• .github/aw/syntax.md

Load these only when relevant:

• .github/aw/experiments.md
• .github/aw/safe-outputs.md

Available Commands

gh aw audit <run-id> --json
gh aw audit <base-run-id> <optimized-run-id>
gh aw logs <workflow-name> --json
gh aw compile <workflow-name>
gh aw status

Start the Conversation

Ask for one of these inputs:

• a workflow run URL (or run ID) to analyze
• a workflow name to review the source
• the guardrail that was exceeded (max-ai-credits, max-daily-ai-credits, max-tool-denials, max-turns / timeout)

Fast Path: Run URL Provided

If the user gives a GitHub Actions run URL:

• Extract the run ID
• Run gh aw audit <run-id> --json
• Inspect agent_usage.aic, agent_usage.input_tokens, agent_usage.output_tokens, agent_usage.cache_read_tokens
• Identify the most expensive phases before asking additional questions

Guardrail-Specific Entry Points

max-ai-credits exceeded

The workflow was stopped because it consumed more AI Credits than the configured per-run budget.

Priority checks:

• Which tool calls dominated token usage? (token-usage.jsonl)
• Is the prompt front-loading large payloads that could be fetched on demand?
• Are there repetitive extraction steps that sub-agents could handle cheaply?
• Does the frontier model handle tasks that a small model could do?

max-daily-ai-credits exceeded

The workflow is being blocked because its 24-hour AI Credits budget is exhausted.

Priority checks:

• What is the run cadence? (scheduled too frequently?)
• Does the workflow use cheap triage before escalating to the frontier model?
• Is batching or caching applicable to reduce run frequency?
• Are there noop early-exits for events that do not require agent action?

max-tool-denials exceeded

The Copilot SDK hit the tool-denial threshold, indicating the prompt attempted actions outside the allowed tool policy.

Priority checks:

• What tool was repeatedly denied? (last denied reason in the failure issue)
• Is the tool missing from the workflow's permissions/firewall config?
• Can the prompt be revised to avoid the denied operation entirely?
• Would a DataOps pre-step satisfy the data need without a tool call?

Timeout / max-turns exceeded

The agent ran out of time or turns before completing the task.

Priority checks:

• Is the task decomposable into smaller, faster sub-tasks?
• Are there long-running tool calls that could be replaced with DataOps pre-steps?
• Is the prompt asking the agent to do too much in one run?
• Can max-turns or timeout-minutes be raised, or should the task be split?

Optimization Analysis Plan

After measuring token usage, produce a prioritized plan:

• Measure — run gh aw audit <run-id> --json and summarize AI Credits and per-call token breakdown
• Diagnose the harness — classify failures across context assembly, tool interaction, generation control, orchestration, memory management, and output processing
• Identify top cost drivers — list the three most expensive phases/tool calls
• Apply quick wins first — DataOps pre-steps, gh-proxy, cli-proxy, prompt trimming
• Sub-agent delegation — identify repetitive per-item loops suitable for small-model workers
• Reuse execution experience — preserve compact task features, configuration deltas, outcomes, costs, and diagnoses in cache-memory when cross-run reuse is useful; apply relevant recurring patterns to similar cases
• Prompt caching — verify stable instructions and reusable experience appear before dynamic content
• Experiment correctness first — add an experiments: entry, compare output quality first, and use metric: "aic" to choose among equivalent-quality variants
• Validate quality — confirm the optimized run produces equivalent safe outputs

Present the plan clearly before making any edits. Confirm with the user before applying changes.

Editing Workflow

• Edit .github/workflows/<workflow-name>.md
• Recompile: gh aw compile <workflow-name>
• Commit both the source and the generated .lock.yml
• Report the estimated savings and link to the PR or commit