💡 Agentic Autofix Fleet Governance & AI Credit Budget Standard

[github-actions[bot]]

Summary

Define the org-wide governance framework for GitHub's new agentic autofix feature (public preview July 2026), which uses Copilot cloud agent to autonomously explore codebases, generate verified CodeQL fixes, and open draft PRs. Establish per-repo opt-in policy, AI credit budget caps, alert-to-autofix routing rules, and human review gates for auto-generated remediation PRs.

Market Signal

GitHub launched agentic autofix for code scanning alerts in public preview on July 10, 2026. Unlike traditional autofix (single-file suggestions), agentic autofix explores relevant files across the codebase, proposes multi-file fixes, reruns CodeQL to verify the fix closes the alert, and opens a draft PR — all autonomously. It draws AI Credits from the org's Copilot allocation. This represents a shift from "found means fixed" to "found means autonomously remediated and verified." Additionally, GitHub Code Quality reached GA on July 20, expanding the scope of findings eligible for automated remediation.

User Signal

The org runs CodeQL across all repos as a required CI check, uses GitHub Advanced Security, and has Copilot licenses — all prerequisites for agentic autofix. The compliance audit workflow (compliance-audit-and-improvement.yml) generates per-finding issues that could be routed to agentic autofix. The org's heavy investment in automated remediation (dev-lead agent, pr-auto-review pipeline) makes this a natural extension. Existing Discussion #761 covers AI-powered code scanning enablement but predates the agentic autofix feature.

Technical Opportunity

The existing compliance audit pipeline already identifies CodeQL alerts and creates issues. Adding agentic autofix routing would close the remediation loop: audit finds alert → assigns to Copilot → agentic autofix generates verified fix → opens draft PR → existing pr-auto-review pipeline reviews it. The canary rollout infrastructure could gate enablement across the fleet (ring0 → ring1 → stable). The pr-limits standard already governs automation PR volume, ensuring agentic autofix PRs don't overwhelm merge throughput.

Assessment

Dimension	Score	Rationale
Feasibility	high	Feature exists in public preview; governance is process/policy, not new tooling
Impact	high	Automates security remediation at scale across all fleet repos
Urgency	high	Feature available NOW; ungoverned enablement risks credit sprawl and unreviewed auto-PRs

Adversarial Review

Strongest objection: The feature is in public preview, not GA. Writing governance for preview features wastes effort if the feature changes or is deprecated. Ungoverned AI credit consumption could burn through the org's Copilot allocation.

Rebuttal: Preview status is precisely why governance should be established now — before ungoverned enablement creates credit sprawl. The standard can be structured as a readiness checklist that evolves with the feature. The org already has precedent for governing preview features (Discussion #484 was created before workflow execution protections reached GA). Credit budget caps are essential precisely because the feature consumes AI Credits, making governance a cost-control necessity, not premature optimization.

Suggested Next Step

Draft standards/agentic-autofix-governance.md defining: (1) per-repo opt-in criteria (must have CodeQL enabled, Copilot license, Code Security), (2) AI credit budget allocation per repo/month, (3) alert severity routing rules (critical/high auto-route, medium/low manual triage), (4) human review requirements for agentic autofix PRs, (5) canary rollout plan for fleet enablement.

[github-actions[bot]]

Weekly Update

What Changed

GitHub Agentic Autofix entered public preview (July 10, 2026). Key capabilities now confirmed:

• Works across both CodeQL and third-party scanning tools — not limited to CodeQL alerts
• Validates fixes by rerunning the original analysis before opening a PR
• Iterates autonomously if the first fix attempt doesn't resolve the alert
• Opens draft pull requests for human review
• Requires GitHub Code Security (or GHAS) + Copilot license with cloud agent enabled

Copilot pre-commit security scanning also launched (July 14, 2026): the Copilot desktop app now scans in-flight code for injection flaws, XSS, and weak cryptography before commit — available to all Copilot subscribers including Free tier.

Industry cost context is critical: Agentic workflows consume 5-30x more tokens per task than standard chatbot queries. Uber's Claude Code adoption went from 32% to 84% of 5,000 engineers between Dec 2025-Mar 2026, burning the entire annual AI budget by April. The AI credit budget aspect of this idea is more urgent than ever.

New related idea created: #999 — Agentic Autofix Adoption & Safety Gate Standard covers the adoption framework and safety gates, complementing this idea's focus on credit budgets.

Sources: GitHub Changelog, Vantage FinOps for AI

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	med	med	→ Requires licensing evaluation; cost metering infrastructure needed
Impact	high	high	→ Now confirmed to work across CodeQL + third-party tools; pre-commit scanning adds another credit consumption vector
Urgency	med	high	↑ Public preview is live; cost explosion risk is proven by industry examples (Uber)

Recommendation

Advance — Agentic Autofix is now available in preview. Prioritize the credit budget and cost metering aspects of this idea alongside the adoption framework in #999. The cost guardrails should be in place before broad autofix adoption.

[github-actions[bot]]

Weekly Update

What Changed

Agentic Autofix reached public preview on July 10, 2026 — this feature has moved from roadmap to deployable. Key developments since this discussion was created:

1. Coverage expanded beyond CodeQL: An editor's note (July 16) clarified agentic autofix now works for all first-party AND third-party code scanning alerts, not just CodeQL. This means the org's zizmor, Snyk, or any SARIF-emitting scanner can benefit.

2. Concrete billing model: During preview, agentic autofix consumes AI Credits from the org's Copilot allocation AND GitHub Actions minutes. Fix generation typically takes 2-4 minutes per alert. This makes budget planning possible.

3. API-driven activation: Fixes can be triggered programmatically via the Update a Code Scanning Alert REST API with assignees set to ["copilot-swe-agent[bot]"]. This enables batch activation via security campaigns — key for fleet-wide remediation of historical alerts.

4. Verified-fix workflow: The agent explores relevant files, proposes a fix, reruns the original analysis to confirm the fix closes the alert, and only then creates a draft PR. Failed verifications don't produce PRs.

5. Requirements clarified: Needs both Code Security (or GHAS) license + Copilot license with Copilot cloud agent enabled. The org already has both.

6. Agent rate-limiting infrastructure now exists: PRs feat: implement issue #638 — [Phase 2] standards/agent-rate-limits.json config + standard doc (single source of truth) #1017, feat: implement issue #639 — [Phase 3] scripts/lib/agent-rate-limit.sh gate + circuit-breaker library (guard-only) #1018, feat: implement issue #641 — [Phase 5] Token-budget circuit breaker: pause at the configured % of the 5-hour Claude window #1020, and feat: implement issue #994 — [Phase 6] Weekly-budget glide-path breaker: pause at 100 − 2×days-until-reset of the 7-day Claude window #1022 (merged Aug 2026) built agent-rate-limits.json and agent-rate-limit.sh — the circuit breaker library that can enforce AI Credit budgets for autofix alongside other agent workloads.

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	med	high	↑ Public preview shipped; rate-limiting infrastructure now exists in this org
Impact	high	high	→ Unchanged; autonomous remediation of security alerts remains high-value
Urgency	med	high	↑ Preview available today; AI Credit budget controls are live; third-party scanner support broadens applicability

Recommendation

Advance. Enable agentic autofix on petry-projects/.github as a canary repo. Set a weekly AI Credit ceiling in agent-rate-limits.json (start conservative — 50 credits/week). Monitor three metrics: (1) credit consumption per fix, (2) fix acceptance rate after human review, (3) impact on PR throughput. After one canary cycle, promote to ring-1 repos if metrics are healthy.