AI Agent Workflows
AI agent workflows are multi-stage pipelines — not a single chat turn. They fail differently than single-shot chat: when every stage depends on the last, end-to-end debugging becomes a whodunit — and models will narrate confident conclusions on top of broken middles.
These posts cover how we bring up, orchestrate, and verify production agent pipelines.
A company-site version of the bring-up write-up also lives on StackGen: How We Debug Multi-Stage AI Agent Workflows. Primary deep-link on this site: How We Debug Multi-Stage AI Agent Workflows.
Part of the series Building an Enterprise AI Agent Platform in Go.
Featured posts
| Post | What you’ll learn |
|---|---|
| From Vague GitHub Issue to PR with Aiden | Board SDLC: Specify→Research→Plan comments, Status hops, optional review PR — and why card-drag ≠ issue webhook |
| How We Debug Multi-Stage AI Agent Workflows | Green one stage at a time; golden gates; score effects not transcripts |
| Evidence-Gated RCA — Prove, Then Narrate | Fixed DAG, structural evals, compound-AI orchestration for SRE RCA |
| Evidence-Based Verification | Don’t trust self-report — check ArgoCD, Datadog, systems of record |
| Is the Task Actually Done? | Goal-scoped completion loops — independent checks, budgets, mutation-safe retries |
| Beyond Confluence Runbooks | Executable GitOps triage vs Confluence narrative — split the corpus |
| Your RCA Agent Needs a Map | Topology, verify-first probes, and learn-from-verdict over runbook-only agents |
| From Demo to Deploy — Failure Modes with Receipts | Umbrella of production-hardening failure modes with receipts |
| The Diary Learning Loop | Organizational learning: propose → human approve → materialize |
| The Hypothesis Ladder | On-call RCA discipline: elimination before narrative |
| AI Agent Root Cause Analysis — Curiosity Before Confidence | Soft prompts vs hard gates for AI RCA; batch validation to stop agent thrash |
| AI Agent Root Cause Analysis — Evidence Discarded After the Lead | Dig found the lead, then abandoned it for peer noise — transcript gates that catch discard |
| AI Agent Loop Detection — Don’t Throw Away the Answer | Preserve the best evidence-backed answer when repetition stops a run |
| How to Steer an AI Agent Mid-Run Without Starting Over | Apply additive human feedback at safe iteration boundaries |
| Single-Agent vs Multi-Agent Orchestration: How to Choose | Decision framework from a fair triage A/B — both shapes have a home |
| Fair Agent Evals: Don’t Compare Until Tools Match | Tool-access parity before planner vs single-agent benchmarks (AppWorld series pt. 1) |
| Agent Orchestration Tax After a Fair Eval | Tokens, iterations, and tool-call overhead when fairness is fixed (pt. 2) |
| AI Agent Eval Failure Modes | Budget, PII placeholders, and self-report vs judge (pt. 3) |
| Stop Spawning Duplicate Workers | Handoff gate sequel: fairness 5/5, lower token tax, harness tuning (pt. 4) |
| Running AppWorld Locally for Aiden Agent Evals | Docker Compose, official MCP HTTP, health gates, and setup pitfalls (pt. 5 — ops prerequisite) |
| How to Evaluate AI Agents: Clarify + Zero-Tool Failures | Unattended eval harness rules: deny clarify, fail zero-tool workers, MCP vs plan layers, policy in TOML (pt. 6) |
| Multi-Agent vs Single-Agent: MCP Tool Tax + pass@k | When planning beats reacting; 453-tool catalog tax; 292-byte handoffs; pass@1 vs pass@3 (pt. 7) |
| Simple vs Plan: When to Use Which | Smoke cohort: plan wins some, ties some; cost tax vs fit — both modes stay in the toolkit (pt. 8) |
Stay in the loop — production notes on AI agents, workflows, and SRE.
Low volume — new posts and curated reading lists. Unsubscribe anytime.
FAQ
What are AI agent workflows?
Multi-stage pipelines where each stage depends on the last — ingest, investigate, verify, notify — run by agents with tools and gates, not a single chat turn.
What is an AI agent workflow vs a single agent turn?
A turn is one plan→tool→answer loop. A workflow chains stages with durable state, skip/loop conditions, and verification so a broken middle cannot hide behind a fluent ending.
How do you debug a multi-stage AI agent workflow?
Bring up one stage at a time against a golden gate — like hardware board bring-up. Green each stage repeatedly before adding the next. Score committed tool calls, not raw transcripts.
What is evidence-gated agent orchestration?
Wrap frontier models in a fixed DAG with structural evals, state merging, and token-aware tool loops. Let Go own pass/fail; let the model narrate only after evidence is committed.
How do you verify agent workflow outcomes in production?
Pull evidence from systems of record — ArgoCD, Datadog, Grafana — instead of trusting self-reported success. Verification gates should be deterministic where possible.