Agentic Development
Use AI agents in enterprise development with context packs, scoped tasks, and review gates.
Key takeaways
- Agentic development succeeds only when agents get a context pack, clear ownership boundaries, and human-reviewable changes.
- A context pack carries goal and acceptance criteria, system map, constraints, commands, and known risks.
- Scope agents to bounded work, keep write ownership clear in parallel runs, and demand changed file paths plus verification results.
- High-risk changes to auth, payments, secrets, legal copy, or deployment config require explicit human approval and cannot bypass CODEOWNERS, CI, or release gates.
- Codex teams encode org constraints in
requirements.toml; Claude teams mirror the same policy in project settings and hooks.
Agentic development works when agents receive the right context, operate inside clear ownership boundaries, and produce changes that humans can review. It fails when agents are treated as unlimited autonomous developers without system knowledge.
Agent Workflow
Context Pack
| Section | Contents |
|---|---|
| Goal | User problem, acceptance criteria, non-goals |
| System map | Relevant apps, packages, APIs, ownership |
| Constraints | Style, security, performance, compatibility |
| Commands | Typecheck, tests, build, local server |
| Risks | Known edge cases and review focus |
Task Scoping
- Assign agents to concrete, bounded work.
- Keep write ownership clear when multiple agents work in parallel.
- Ask for changed file paths and verification results.
- Review generated code with the same standards as human code.
- Save recurring workflows as skills or templates only after they prove useful.
Governance
Agent outputs must pass normal CI and code review. High-risk changes involving auth, payments, secrets, legal copy, or deployment configuration require explicit human approval.
Multi-Agent Policy
| Policy | Requirement |
|---|---|
| Context | Keep project instructions, context packs, and skills reviewed like code |
| Isolation | Use branches or Git worktrees for long-running implementation tasks |
| Permissions | Start with least-privilege file, shell, network, and MCP access |
| Verification | Require changed files and command results in the final agent report |
| Merge | Never let agent output bypass CODEOWNERS, CI, or release gates |
For Codex-based teams, place organization-wide constraints in managed configuration or
requirements.toml; for Claude-based teams, mirror the same policy in project settings and hooks.