AI Architecture Bootcamp Archives - Page 8 of 10

Day 57: Coding Agents That Only Emit Intents

Coding agents emit typed intents; an AST/codemod apply layer patches files so LLMs never own raw bytes.

Capstone spec template: users, SLO, tools, eval, rollback — pick one production problem and freeze scope.

Build vs buy the agent runtime: Bedrock Agents vs custom loop vs IDE hooks — decide with constraints, not hype.

Bedrock supervisor agents with action groups that typecheck against handlers — fail CI on OpenAPI drift.

Detect split-brain plans and duplicated PRs with single-active-plan locks, implementer leases, and PR fingerprints.

Hire and topology for AI platform: platform vs embedded ownership, and who owns eval.

When the agent goes wrong: kill switches, replayable traces, and customer-safe comms.

Least privilege per role: Critic cannot shell; Implementer cannot touch prod IAM — enforce with tool matrices and IAM roles.

Cap rounds, tokens, and wall-clock; detect oscillation; escalate to a human after N critic failures instead of overnight ping-pong.

Prompt change management means versioned Prompt Management — no hot-edits in production.