You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Epic] Make the Code Review Engineer a real, differentiated employee #10
AgentOS markets a marketplace of specialized AI employees, but exploration showed the employee types aren't actually differentiated: Code Review Engineer, Customer Support, and Secretary run the identical container with the identical code and skill set — the role is just a string env var. The template fields that would make a role distinct (system_prompt, allowed_actions) are parsed but ignored, and the "5-layer enforcement moat" from the docs is entirely unbuilt.
Decision: go deep on one role rather than add more undifferentiated shells. Turn the Code Review Engineer into a genuinely real employee — behavior driven by its template, boundaries provably enforced, autonomous, and memory-backed. The engine built here (template → runtime) is also what would make any future breadth or marketplace play cheap.
The Code Review Engineer is now a differentiated, trust-moated, autonomous, memory-backed employee end-to-end: hire → subscribe to a repo → watcher dispatches a review on the next open PR → review posted on GitHub with the per-agent token, action policy enforced, every call audit-logged, dedup against reviewed_prs.
Remaining depth work tracked separately: #33 (pipe the diff into the review prompt so it references actual code), #23 (memory compaction), #17 (bring Customer Support & Secretary to A/B/D parity).
Activity
michaelzwang13 commentedon May 25, 2026
All four phases shipped:
The Code Review Engineer is now a differentiated, trust-moated, autonomous, memory-backed employee end-to-end: hire → subscribe to a repo → watcher dispatches a review on the next open PR → review posted on GitHub with the per-agent token, action policy enforced, every call audit-logged, dedup against reviewed_prs.
Remaining depth work tracked separately: #33 (pipe the diff into the review prompt so it references actual code), #23 (memory compaction), #17 (bring Customer Support & Secretary to A/B/D parity).