Define agent boundaries, integrate into GitHub workflows, and establish observability
Domain Weight
Prepare Agent Architecture and SDLC Processes accounts for 17% of the GH-600 exam.
Prerequisites Before Deployment
Inputs — what data, events, or context the agent receives (issues, PRs, logs, specs)
Outputs — what artifacts or actions the agent produces (PRs, comments, test results, plans)
Success criteria — measurable conditions that define correct agent behavior
Without all three, the team cannot objectively validate whether the agent succeeds or fails
Plan → Validate → Execute
Key Pattern
Separate agent planning (reasoning) from execution (action). The plan is reviewed by humans or automated policy checks before any irreversible action is taken.
Structured plan output: JSON/YAML written to PR descriptions or CI workflow summaries
Pre-execution approval gates for high-risk or untested agents
Diff comparison between planned steps and actual tool calls to detect divergence
Anti-pattern: combining planning and execution in one indivisible step
SDLC Integration Patterns
Integration Point
Correct Pattern
Anti-Pattern
Code changes
Agent opens PR; human approves before merge
Agent pushes directly to protected main branch
PR review
Agent posts structured review comments; human merges
Agent auto-approves and merges its own PRs
CI workflows
Agent invoked as a workflow step with scoped secrets
Agent runs with org-wide credentials 'just in case'
Permissions
Least-privilege scoped to task repos/branches
Agent self-modifies permissions during a run
Audit trail
Durable logs in GitHub Actions summaries and PRs
In-memory-only logs with no persistent record
Graduated Autonomy Model
Phase
Autonomy Level
Human Role
Phase 1
Supervised — read-only or every action approved
Must approve every action
Phase 2
Semi-autonomous — low-risk writes autonomous
Approves high-risk actions only
Phase 3
Fully autonomous for proven low-risk operations
Reviews outputs post-hoc
Inspectable Agent Artifacts
Plans: structured reasoning traces in PR descriptions or workflow summaries
Diffs: code/config changes surfaced as standard GitHub PR diffs
Decision logs: step-by-step reasoning accessible in Actions output
Action logs: tool calls with parameters, timestamps, and results
Completion status: explicit success/failure with reason — no silent failures
