AI Agent Workflow Collection

Use when tracking a slowly-varying latent quantity from noisy measurements with quantifiable uncertainty — a drifting metric, capacity, latency trend, or confidence. Prefer a Kalman-style adaptive estimator over fixed moving averages or static thresholds.

Use when an autonomous loop needs explicit limits on attempts, time, cost, tool calls, verification depth, or risk, and a clear stop, continue, yield, or escalate decision.

Use when adopting a shared module, library, chart, or service, or doing any consolidation, migration, or refactor that supersedes existing code. The change that adds the replacement must REMOVE what it replaced, in the same atomic change.

Use when an agent operates a UI on the user's behalf through conversation, navigating, filling forms, updating live views, with real-time bi-directional state sync. Activates on conversational app control, agent drives the front end, agentic UX, or choosing an agent-to-app protocol.

Emit OpenTelemetry trajectory spans (tool-call, latency, token usage) as an agent works — vendor-neutral, via otel-cli wrapping shell tool-calls, exported to any OTLP backend.

Use when implementing code from a spec through a bounded build-test-fix loop with explicit acceptance criteria, automated checks, live verification when needed, and clear stop conditions.

Use when onboarding, scaling, decommissioning, or auditing autonomous agents (or any "software employee") in a fleet. Model each agent as a reusable infrastructure-as-code module whose inputs are its facets, so its whole configuration is declarative, reviewable, and reproducible.

Use when exploring or communicating a UI layout before writing code. Sketch the screen as a text mockup — boxes, fields, labels — and offer a few distinct variations, so the layout decision is made cheaply in a PR or chat before anyone builds it.

Build client-facing HTML email that actually renders — correct multipart MIME (HTML + plain-text fallback ALWAYS), inline styles, and verify by rendering the received message.

Use when deciding whether a capability should be a SKILL or an MCP SERVER, and how to author either well. The axis is knowledge-vs-capability: would the agent READ this to know how, or CALL this to do it?

Use when a loop or watchdog SSHes across many hosts to probe, enumerate, or act. Force non-interactive SSH with timeouts so one un-authable host cannot hang the loop, and check a heartbeat so a stuck-but-alive process is detected.

Use when a request bakes in an assumption that may be wrong — a leading question, an already-chosen approach, a framing. Interrogate the premise before answering as asked; if the question itself is flawed, say so and answer the better question, with reasoning.

Use when starting code exploration or implementation in a repository where CodeGraph may be available. Ensure CodeGraph is installed/initialized when authorized, then prefer CodeGraph queries for semantic code context.

Use when starting or resuming code work in a repository worktree. Initialize or sync CodeGraph, keep its cache uncommitted, and prefer CodeGraph for code navigation before grep or broad file reads.

Use when a Kubernetes Job, CronJob, or Pod is pinned to a specific node (via podAffinity, nodeSelector, or nodeAffinity) to share a node or an RWO volume. It must also carry that node's tolerations or it is permanently unschedulable.

Use when code review needs enforceable limits on branching, nesting, function size, or parameter count. Ratchet budgets through standard linters.

Use when installing, repairing, or verifying an MCP or other authenticated connector, especially when configuration, authentication, quota, and real artifact access can be mistaken for one another.

Use when patching code you do not fully control or that moves between runs. Anchor edits on unique surrounding context and verify by behavior, never by line number or a grep that only confirms the text is present somewhere.

Use when an agent should treat GitHub issues as an ongoing work queue, including triage, decomposition of oversized issues into smaller child issues, PR-based implementation, verification, and evidence-backed issue closure.

Use when a workload (agent, sandbox, container, CI job) must call an authenticated upstream (LLM/provider/cloud API) but should not hold the credential itself. Keep the secret in a proxy that injects it, so a compromised workload cannot exfiltrate the key.

Use when deciding what a service or worker should report and how loudly — alerts, progress updates, logs, notifications. Tier output by criticality so the important signal is not buried in routine chatter.

Use when building integrations or replacing synthetic demos with real data sources. Requires source contracts, idempotent ingestion, validation, observability, and fixture-backed tests.

Use for any change to a system's behavior, anywhere. Default to declaring desired state in version-controlled source that tooling reconciles, not making a live mutation. The test for any change is "where is the declared source, and what reconciles it?

Use when disk usage is high, df and du disagree, or an agent must diagnose full filesystems without deleting valuable work.