July 1, 2026 GuideADR

Setting up an agent production engineer

A production-ticket runner that starts from evidence, routes read-only specialists, and stops before irreversible actions.

The agent begins with Linear and current production evidence, not a source-code search. It produces one evidence bundle, then chooses one outcome: stop, escalate, no-code, vendor or infra handoff, or PR.

This is the core shape of the PRODE suite: a coordinator inside a durable runner, supported by read-only specialists and constrained by explicit human approval gates.

The system is useful because production tickets rarely fail from lack of code access. They fail from missing context, stale assumptions, unclear ownership, insufficient evidence, and unsafe escalation paths.

Humans own irreversible actions: production writes, deploy-risk changes, PR merge, Linear Done or Cancelled transitions, broad telemetry suppression, and sensitive data boundaries.

Concepts first

The terminology matters because it prevents the agent from becoming one large unbounded prompt. Each concept has one job.

Runner

The outer control plane for leases, retries, state, idempotency, approval waits, and resume-after-crash behavior.

Coordinator

The main agent process. It owns the ticket decision, routes specialists, classifies the work, and writes handoffs.

Specialists

Read-only evidence workers for Sentry, AWS, database, vendor, support, release, reproduction, code path, and review.

Evidence bundle

The go/no-go artifact: sources, facts, confidence, gaps, classification, and approval state.

PRODE Wiki

Operational memory. It can suggest where to look; it cannot prove the current incident.

Stop state

A deliberate safe outcome when evidence, confidence, ownership, or approval is not sufficient.

One run, end to end

The run is intentionally linear until the evidence bundle exists. Parallelism happens inside specialist gathering, not inside the final decision.

Lease the ticket

The runner acquires linear:{ticket_id} so one automation owns the work.

Read Linear

The coordinator fetches ticket text, labels, comments, attachments, Sentry URLs, request IDs, and timestamps.

Inventory tools

It records which MCPs, connectors, CLIs, skills, and blocked capabilities are available.

Read memory

Relevant PRODE Wiki pages may guide queries and negative controls, but they do not count as evidence.

Route specialists

Specialists gather current facts from Sentry, AWS, DB, vendor, release, support, reproduction, and code paths.

Bundle evidence

The coordinator records sources, findings, confidence, gaps, classification, and approval state.

Decide

Below 70% confidence, missing evidence, or pending approval stops the run. A narrow code fix may proceed.

Fix and hand off

After the evidence gate, the agent reads code, patches, verifies, opens a PR, and updates Linear.

Monitor after merge

After a human merge signal, the runner checks recurrence and proposes redacted learnings.

Setup

The suite works only if the repo points agents at one authoritative workflow and the runtime records what happened outside the prompt.

TriggerWebhook, scheduled scan, manual dispatch, or human command. Every trigger resolves to linear:{ticket_id}.
Entry point.agents/skills/prode-triage/SKILL.md is the normative workflow. Repo instructions point to it rather than duplicating rules.
SecretsUse secure local or platform secret storage for Linear, GitHub, Sentry, AWS, and read-replica credentials.
ArtifactsLinear holds summaries; the runner store holds state and pointers; external storage holds redacted bundles; git holds code, PRs, reviewed docs, and candidate learnings.
Canonical filesInstall the coordinator skill, specialist reference, automation runner reference, redaction lint, learning inbox, and PRODE Wiki paths.

Runner contract

The runner persists state; the coordinator makes decisions; specialists gather evidence; humans approve irreversible actions.

Lease

Acquire, heartbeat, expire, and resume. Never race another active run.

Idempotency

Find existing comments, branch, PR, specialist outputs, and artifacts before creating new ones.

Approval wait

Pause until approval matches the exact action, approver, timestamp, and expiry.

Evidence and confidence

The evidence bundle is the transition point between investigation and implementation. The agent may implement only when required specialists are complete or explicitly skipped, confidence is at least 70%, and approvals are satisfied.

Raises confidenceSentry event, CloudWatch logs, release timing, DB state, or reproduction all point to the same mechanism.
Lowers confidenceOnly one weak source exists, logs are missing, request IDs are absent, or specialists disagree.
Blocks automationSpecialists conflict by more than 20 points, approval is required, evidence is missing, or the next action belongs to a vendor, infra owner, or human incident process.

Safety

The agent is powerful at reading and conservative at mutating. Production DB access is read-only and narrow. Sentry, AWS, vendor, and GitHub mutation is blocked or approval-gated. Linear receives summaries, not raw production payloads.

Allowed

Request IDs, Sentry issue IDs, releases, versions, log groups, and redacted source names.

Never include

Secrets, tokens, contact details, payment identifiers, KYC fields, copied DB rows, or full payloads.

Stop

Any production write, lock, migration, backfill, broad suppression, merge, or Done/Cancelled transition waits for humans.

Rollout

Roll out the suite by making one skill authoritative, adding the safety substrate, then moving from assisted investigations to automated PR handoff.

Canonical suite

Keep prode-triage authoritative and wrap or remove old split skills.

Safety substrate

Add learning inbox, wiki ignore rules, redaction lint, and CI checks.

Runner

Implement durable state, leases, approvals, evidence bundles, PR handoff, and post-merge monitoring.

Wiki

Promote only reviewed, redacted, source-backed learnings into durable operational memory.

Keep reading

Agent Production Engineer Setup Guide

The comprehensive standard ADR page with diagrams, acceptance criteria, and implementation details.

Guide
ADR-002: PRODE Engineering Agent Suite

The underlying decision record for the suite, specialist routing, and automation runner.

ADR