One code change left its rationale scattered across five systems
Software delivery involved architecture, information security, compliance, operations, and release management in addition to engineering. In an environment built around separation of duties, introducing AI exposed a traceability gap.
Answering “why did the AI suggest this?” required reconstruction
Developer-agent conversations, source changes, pull requests, security questions, architecture discussions, and deployment approvals lived in separate systems. Weeks later, reviewers had to reconstruct the evidence manually.
Execution control alone could not explain collaboration context
Patty Code handled development execution and PCCP controlled models, data, and tools. But existing collaboration systems still behaved as if the AI did not exist. Crew was introduced to connect human and agent activity in the same project context.
Crew was scoped to AI-involved projects, not company-wide messaging
The 12-week rollout covered 60 developers, 12 security, architecture, and operational reviewers, 15 repositories, and three application teams. Crew was narrowed to the authoritative workspace for AI-involved engineering rather than a replacement for ordinary messaging.
Implementation
Execution and authority remained separate
While Patty Code investigated and executed changes, PCCP governed approved models, source access, sensitive information, package installation, environment boundaries, and role permissions.
AI became an identifiable participant, not pasted text
In projects such as authentication modernization, developers, the coding agent, security reviewers, and architecture reviewers worked in one space. The agent retained its own activity and context, allowing reviewers to inspect both the change and the surrounding collaboration.
Observed change
- 2–3시간 → 20–30분
- Evidence preparation for significant changesObserved across many participating-team changes
- 약 3일 → 약 2일
- Median AI-assisted review cycleCustomer observation during controlled rollout
- 1개의 연결된 맥락
- Engineering, AI, security, and architecture activityAuthoritative record for AI-involved work, not a universal tool replacement
The first Crew scope was too broad
Attempting to move ordinary team communication created resistance. Adoption improved when Crew was limited to engineering work in which agents participated.
PCCP explained authorization; Crew explained the work context
Security and architecture valued Crew as much as developers did. Pilot was not yet necessary because agents were still initiated and supervised directly by people.
Evidence and disclosure
Customer-reported and pilot-observed. Timing figures are approximate ranges for meaningful AI-assisted change reviews in participating teams.
The institution, project names, and identifying system details are withheld. Interview notes are reconstructed without presenting direct quotations.