Internal developer tooling · Live workflow
08 / 12
Slack AI Engineering Agent
An operational Slack-based agent workflow that turns high-confidence Sentry noise into validated draft pull requests for human review.
- Role
- Feature Engineer / Product Engineer · 2026
- Problem
- The agent needed enough authority to remove repetitive operational work without gaining arbitrary permission to change or merge application code.
- Strategy
- The workflow converts only high-confidence alerts into a typed action, deduplicates them with a stable fingerprint, applies a deterministic patch inside an isolated branch, validates the allowed diff and targeted test, then opens a draft pull request instead of merging.
- Outcome
- The workflow operated in Slack and produced several draft pull requests for human review.
Operational internal workflow · Private source
01
Problem definition
Why this problem had to be solved
Recurring browser-wallet extension noise reached engineering teams through Sentry and repeatedly required the same triage, filter change, regression test, and pull-request preparation.
What users and the system were facing
Known extension noise repeatedly consumed manual Sentry triage and filter-PR work across product frontends.
A Slack alert could provide context, but it could not safely grant an agent open-ended permission to modify repositories.
Generated filters had to suppress only extension-owned noise without hiding application errors or bypassing human review.
Frontend ownership
Within a larger team-built agent codebase, I owned the Sentry automation subsystem from its RFC, subscription action and persistence model through guarded patching, validation, GitHub handoff, and Slack feedback.
RFC, event model, and explicit subscription actions
Sentry parsing, confidence gates, fingerprinting, and deduplication
Deterministic filter patching, regression tests, and diff guards
GitHub preflight, draft-PR creation, audit records, and Slack feedback
02
Solution strategy
A system strategy carried into production
The workflow converts only high-confidence alerts into a typed action, deduplicates them with a stable fingerprint, applies a deterministic patch inside an isolated branch, validates the allowed diff and targeted test, then opens a draft pull request instead of merging.
System model
A Slack-delivered Sentry alert is classified and deduplicated before guarded patching, targeted validation, draft-PR creation, and human review.
Implementation decisions
Kept background subscriptions read-only and routed only an explicit opt-in action into the code-changing executor.
Required extension-owned stack frames and generated both positive and false-positive regression tests so application errors remained visible.
Limited repositories, files, execution frequency, and branch scope while masking secrets and cleaning isolated worktrees.
Produced draft-only pull requests and returned success, duplicate, and failure states to Slack for human review.
03
Outcome
What changed—and what I learned
The workflow operated in Slack and produced several draft pull requests for human review.
It materially reduced recurring manual triage and filter-PR preparation without relying on an unverified time-saved figure.
It preserved the existing permission boundary with allowlists, rate limits, audit records, and draft-only delivery.
Learning
Useful AI engineering automation comes from narrow authority, deterministic validation, visible failure states, and a deliberate human handoff—not from maximizing autonomy.
Today
I would promote the feature tests into the default CI gate and track alert-to-draft conversion, review acceptance, false positives, and rollback outcomes as explicit operating metrics.
Built with
Public links
Architecture-level explanation only; private source and internal payloads are intentionally not shown.
Work together