Agentic AI operating guide

Use when a team has an AI idea but has not yet defined the consequential job it should improve.

Write a bounded agentic AI brief for [TEAM]. Define the user or operational decision, the current workflow, the desired outcome, the people affected, and the smallest useful improvement. Separate known evidence, assumptions, and missing information. Recommend whether the problem needs a deterministic automation, a single AI agent, or coordinated agents, and explain why without inventing metrics.

Goal:
[PASTE GOAL]
Evidence:
[PASTE EVIDENCE]
Constraints:
[PASTE CONSTRAINTS]

Output: A decision-ready brief with a bounded recommendation.

Guardrails: Do not assume autonomy is the answer · Separate facts from assumptions · Name the human owner

Use when a team is deciding how much reasoning and autonomy a process actually needs.

Compare three implementation patterns for this process: deterministic workflow, single AI agent, and coordinated agents. Score each against variability, tool use, latency, data sensitivity, failure cost, observability, human review, and maintenance. Return a recommendation, the smallest pilot, and the conditions that would justify moving to a more autonomous pattern. Keep scores qualitative unless evidence supports numbers.

Process:
[PASTE PROCESS]
Systems and data:
[PASTE SYSTEMS]
Risk tolerance:
[PASTE RISK CONTEXT]

Output: An architecture choice with a staged pilot and escalation criteria.

Guardrails: Do not equate more agents with more value · Keep high-consequence writes human-approved

Use when a workflow needs a shared model of what the system observes, decides, does, and learns.

Map this workflow into five observable stages: perception, reasoning, planning, action, and reflection. For each stage list inputs, approved context, expected artifact, tools, owner, timeout, failure state, and hand-off. Do not expose private chain-of-thought; describe concise rationale and evidence a reviewer can inspect. Finish with the event schema needed to trace one run end to end.

Workflow:
[PASTE WORKFLOW]
Available tools:
[PASTE TOOLS]
Human roles:
[PASTE ROLES]

Output: A state map and traceable event schema.

Guardrails: Use observable artifacts instead of hidden reasoning · Every action has an owner and failure path

Use before connecting an agentic system to internal data, APIs, or customer-facing actions.

Design the context and tool boundary for this agentic workflow. Classify each data source as approved, conditional, or prohibited. Classify each tool as read, draft, or write, and define the permission and approval required. Specify freshness, provenance, retention, rate limits, and the safe behavior when a source or tool is unavailable. Return a compact policy table and the minimum tool surface for the first release.

Workflow:
[PASTE WORKFLOW]
Data sources:
[PASTE SOURCES]
Tools:
[PASTE TOOLS]

Output: A least-privilege context and tool policy.

Guardrails: Default to least privilege · Separate read, draft, and write · Do not hide provenance

Use when an agent may recommend, draft, or execute actions with real-world consequences.

Create an autonomy policy for this agentic system. Define what it may observe, suggest, draft, execute, and never do. Add approval thresholds, escalation triggers, identity and access requirements, audit events, incident ownership, and a kill switch. For each consequential action, specify the human review evidence required before execution.

System:
[PASTE SYSTEM]
Actions:
[PASTE ACTIONS]
Risk and compliance context:
[PASTE CONTEXT]

Output: A reviewable autonomy, escalation, and incident policy.

Guardrails: Consequential writes require explicit approval · Name the approver and recovery owner · Provide a kill switch

Use before launch or when adding a new tool, data source, model, or hand-off.

Threat-model this agentic workflow. Cover prompt injection, tool misuse, data leakage, privilege escalation, untrusted content, model drift, supply-chain risk, denial of service, and unsafe hand-offs. For each threat state the asset, attack path, preventive control, detection signal, response owner, and residual risk. End with launch blockers and monitored follow-ups.

Workflow and tools:
[PASTE WORKFLOW]
Data and identities:
[PASTE DATA]
Current controls:
[PASTE CONTROLS]

Output: A threat model with controls, signals, and owners.

Guardrails: Treat retrieved content as untrusted input · Do not rely on a model instruction alone as a security boundary

Use before production launch and after changing prompts, models, tools, or policies.

Build an evaluation set for this agentic workflow. Include representative cases, ambiguous requests, missing-context cases, adversarial or prompt-injection cases, tool failures, and human escalation cases. For every case define expected observable behavior, acceptable variation, required evidence, unacceptable action, reviewer rubric, and telemetry. Recommend release blockers and monitored follow-ups.

Agent job:
[PASTE JOB]
Tools and permissions:
[PASTE TOOLS]
Known failures:
[PASTE FAILURES]

Output: A release gate and evaluation rubric grounded in observable behavior.

Guardrails: Test unsafe and ambiguous inputs · Evaluate actions and evidence, not hidden reasoning

Use when an agentic workflow is live and the team needs to improve quality, cost, and trust.

Design an operating dashboard for this agentic workflow. Define leading and lagging measures for task success, human override, escalation, tool errors, latency, cost, data freshness, safety incidents, and user trust. For every measure specify event source, calculation, segment, alert threshold, owner, and action. Add a weekly review agenda and a rule for when to pause or roll back autonomy.

Workflow:
[PASTE WORKFLOW]
Business outcome:
[PASTE OUTCOME]
Telemetry available:
[PASTE TELEMETRY]

Output: A measurement and review system for continuous improvement.

Guardrails: Do not optimise proxy metrics without outcome checks · Include a pause and rollback rule

Use when support or operations work is repetitive but still needs human accountability.

Scope a supervised agentic pilot for this customer operations workflow. Define the customer job, intake signals, approved knowledge, draft response or action, escalation conditions, human approval, recovery path, and success measures. Include one workflow the system should not automate and a two-week learning plan. Keep customer-facing claims traceable to approved sources.

Current workflow:
[PASTE WORKFLOW]
Knowledge sources:
[PASTE SOURCES]
Support policy:
[PASTE POLICY]

Output: A bounded customer-operations pilot with review and learning criteria.

Guardrails: Escalate sensitive or ambiguous cases · Do not invent policy or customer history

Use when search, content, or paid-media work needs speed without handing publishing or spend to an agent.

Design a supervised growth agent for [WORKFLOW]. It may research, cluster, draft, or recommend, but it must not publish, change spend, or make unsupported claims. Define sources, tools, output schema, evidence requirements, human approval, experiment measure, and rollback path. Return a launch brief with three representative cases, two edge cases, and one prohibited action.

Workflow:
[PASTE WORKFLOW]
Business goal:
[PASTE GOAL]
Approved sources and tools:
[PASTE SOURCES AND TOOLS]

Output: An approval-ready growth agent brief with evidence and rollback.

Guardrails: No auto-publishing or spend changes · Keep recommendations separate from facts · Require source evidence