02 Case study
Sentinel
Sentinel is an always-on quality platform. Every 15 minutes, a scheduler heartbeat runs three specialized agents against registered repos: code review, docs-sync, and agentic QA. Each emits structured reports with machine-parsed verdicts. No human presses a button.
- Client
- SimbaStack, internal
- Sector
- Developer infrastructure / autonomous agents
- Year
- 2026
- Role
- Architect and builder
- Engagement
- Internal platform
The problem
Quality work is the first thing a small team drops under pressure. Code review queues grow, docs rot, and regression testing happens the night before release, if at all.
Existing CI catches what you already wrote tests for. It does not notice that check-in silently stopped updating room state, because no assertion ever knew to look.
The approach
Sentinel treats quality as infrastructure. Three agents run on a 15-minute heartbeat: a code reviewer, a docs-sync agent, and an agentic QA tester. Each produces a structured report with a machine-parsed verdict, so results feed dashboards instead of inboxes.
The QA agent drives a real browser. An LLM reads the live DOM and plans each action while Playwright executes it. A deeper flow engine goes further: it recons the repo's routes, APIs, and schema, derives the critical business flows on its own, then asserts UI-versus-backend consistency through API calls.
A vision pass scores the UI against a WCAG and heuristics rubric using a multimodal model. The docs-sync agent edits Markdown only, inside isolated git worktrees, behind a hardened allow-list guard. It never auto-merges.
Architecture
A scheduler heartbeat fans out to sandboxed agents. Every write path is guarded; every result is structured.
- 15-minute heartbeat
- Repo registry
- Single-flight per repo
- Code review
- Docs-sync (worktree-guarded)
- Agentic QA (browser-driving)
- Route / API / schema recon
- Derived business flows
- UI vs backend assertions
- Allow-list write guard
- No auto-merge
- Adversarial security audit, findings fixed
- Structured reports
- Machine-parsed verdicts
- Vision-based UI/UX scores
The hard parts
Agents that touch repos must be caged
The docs agent writes files, which makes it the most dangerous component in the system. It operates in an isolated worktree behind an allow-list guard that permits Markdown edits and nothing else. The design was then hardened through an adversarial multi-agent security audit, and every high-severity finding was fixed.
Cost decides whether it survives
An agent fleet nobody can afford gets turned off. Careful model selection and prompt design hold a full QA run at roughly three cents, cheap enough to run around the clock without a budget conversation.
Deriving flows nobody wrote down
The flow engine cannot rely on existing test suites. It reads routes, API surfaces, and schema to infer what the product is supposed to do, then tests that. This is what let it catch bugs no human had written an assertion for.
Outcome
Sentinel caught real reservation-state bugs in KaribuKit before release: stale check-ins and phantom room availability that surface-level passes missed.
It is quality infrastructure in the literal sense: always on, structurally cheap, and incapable of merging its own work. The same pattern now extends to any repo the studio registers.