AI Agent & Automation Workflows

Invoke when iterating toward a fix/optimization/answer against a live system and tempted to guess from sparse signals — an observe→analyze→act loop over instrumentation + a scalar metric + one-variable bounded experiments + a keep-or-revert gate + an append-only ledger; covers instrument-before-you-hypothesize, the stall escape-hatch, and the guess-and-rerun / stale-context / "flaky" / claim-without-evidence anti-patterns that make iteration spin.

Invoke when a system is slow, saturated, regressed, or near capacity and you are tempted to guess the bottleneck or reach for a random tool — Brendan Gregg's systems-performance methodology. Covers the USE method (check utilization, saturation, errors for EVERY resource), the 60-second triage checklist, TSA / thread-state, workload-characterization / drill-down / latency methods, the per-resource Linux tool checklist (vmstat, mpstat, iostat, sar, pidstat, perf, bpftrace/bcc), and GPU performance counters (nvidia-smi, dmon/pmon, DCGM, XID/ECC/NVLink, Nsight). Use for capacity, latency, throughput, "where is the bottleneck", and CPU/memory/disk/network/GPU questions.

Invoke when you have a bug report or issue and need to confirm it's REAL before anyone fixes it — turn the report into a precise claim, faithfully recreate the reporter's conditions (the version they hit, config, data, platform, timing), build a minimal reliable reproduction, quantify how often it fires, and return a verdict: CONFIRMED / PLAUSIBLE / BUSTED. Use when triaging an issue, when "can't reproduce" / "works on my machine", when a failure is intermittent or flaky, or before assigning, fixing, or closing a bug. The reproduce-before-you-believe step upstream of root-causing.

Invoke when comparing models or prompts and you want evidence, not a vibe check — choosing a model for a task (opus vs sonnet vs haiku; Claude vs Codex vs Gemini), A/B testing prompt variants, running a model/prompt bake-off or eval, or confirming a prompt change actually helped rather than got lucky. Runs a controlled A/B/n experiment: one variable, the same test cases through every arm, N replicates for stochastic noise, a blinded judge, a pre-registered metric and decision rule, and an honest "is the gap beyond noise?" — not "B felt better." Starts with a Claude adapter; extensible to Codex/Gemini.

Invoke when orchestrating a FLEET of parallel long-running agent tasks via a priority-queue manager over git worktrees — agent queue + task list + dispatched agents + a self-pacing heartbeat loop; covers dispatch, worktree isolation, the manager-only merge protocol, and the resource/orphan/determinism anti-patterns that make fleets fail.