Argus: Cross-Repo Regression QA Agent
Caplia's in-house QA agent. One Slack message dispatches browser-automation workers across every app and posts a per-flow PASS, FAIL or BLOCKED digest back in the same thread.
“Pre-release smoke testing went from hours of manual clicking to one Slack message.”
Problem
Before Argus, pre-release smoke testing meant the team manually clicking through user journeys for hours before any merge to production. Regressions across the founder app, the investor app and shared database migrations were caught late, or by users.
Solution
You @-mention Argus in Slack, name a flow or an area, and it handles the rest: dispatches a GitHub Actions run, fans out browser-automation workers across the apps, aggregates the artefacts and posts a per-flow digest in the thread. Keep replying to ask why a flow failed, request a re-run or pull the diff that triggered the regression. No CLI, no dashboards. Built in about two weeks and in daily use since.
Overview
A Slack edge function on Supabase hosts a Claude tool-use loop with seven tools: catalogue lookup, cross-repo diff, run dispatch, status, summary, failure trace and cancel. The orchestrator runs in GitHub Actions, where an LLM router picks which flows to run from the diff between prod and main. Workers run on the Claude Agent SDK with chrome-devtools-mcp against a real browser, real auth and real network, spawned in parallel, one per app. Thread state lives in Postgres with single-flight locks so concurrent messages in one thread never collide, and hallucination is guarded at two layers: prompt rules for the model plus deterministic post-run checks in code. Not every layer needs the frontier model. The workers do rote DOM and network work, so they moved to a cheaper Anthropic-compatible model while Claude kept the orchestrator and the Slack-facing brain, roughly 10 times cheaper per run with no measurable drop in test quality. Argus now gates every production deploy: a 672-flow regression catalogue across five repos, a full dispatch in about 90 minutes for single-digit pounds, a narrow run by area in about five minutes for under a pound.
Ask my agent about Argus
It knows the problem, the build and Daniel's part in it, and can book a call if you want to go deeper.
Ask the agent