Agents propose, people and policy accept

The boundary in this pattern is simple: an agent may propose work (a plan, a branch, a pull request), but controls outside the agent decide whether it is accepted. A confident agent cannot skip a gate, because the gate is enforced by the platform, not by instructions the agent is trusted to follow.¹

The same boundary in three sources

Source	Where the boundary sits
GH-600 course, module 2	Required checks, CODEOWNERS review, branch protection, and environment approvals decide what merges and deploys.1
AI-native SDLC playbook	Agents propose changes only through pull requests and branch protection; the production gate stays human-controlled.2
Warp’s self-improving agents	Improvements to an agent’s own skill file arrive as pull requests the team reviews and can revert.3

Making it enforceable on GitHub

The GH-600 module turns the boundary into three mechanisms:¹

1. A pull request template requiring goal, scope, steps, verifiable success criteria, risks, and a rollback plan.
2. A required status check (for example a “Plan Gate” workflow) that fails when the plan is missing.
3. CODEOWNERS, so changes under paths such as /security/, /.github/workflows/, or /infra/ need the owners’ sign-off.

Plan first, or plan with the code

	Plan-first PR	Plan and execution in one PR
Plan visible	Before any code exists	Alongside the first commits
Human validation	Before code is written	Before merge
Suited to	High-risk, hard-to-reverse changes	Low or medium risk, easily reversed

Both are safe when GitHub protections are configured; the only variable is when code may exist relative to approval. Planning agents should get read-only tools, with a deliberate handoff to an implementation agent.¹

Version control also supplies the audit log, approval gate, and rollback that a bespoke agent-memory system would have to build (analysis in the Warp note).³ The SDLC playbook makes the same point: Git history records what was asked, what the agent produced, which policy applied, and who approved it.²

Related

• Risk-based autonomy: how much each kind of change is allowed to do before a person looks.
• Delegation contract: what the proposal must contain.
• Source: GitHub Certified: Agentic AI Developer
• Source: The AI-Native SDLC Playbook
• Source: How Warp Builds Self-Improving Agents on Claude

Footnotes

1. GitHub Certified: Agentic AI Developer (study notes) ↩ ↩² ↩³ ↩⁴

2. The AI-Native SDLC Playbook ↩ ↩²

3. How Warp Builds Self-Improving Agents on Claude ↩ ↩²