An access-request agent that checks role policy and separation-of-duties, auto-provisions low-risk access, and escalates privileged or sensitive requests.
ClaudeLangGraphOpenAIAdvanced
–
60+ blueprints · open & free
60+ blueprints across 30 categories. Read the design, copy the prompt, download the kit. Free.
{
"name": "Refund Resolution Agent",
"trust_level": "A3",
"dna_pattern": "retrieve · reason · propose · gate · act",
"worst_case_action": "Issue a refund to the wrong account",
"authority_boundary": {
"can_modify_records": true,
"can_delete_records": false,
"can_send_messages": true,
"can_spend_money": true
},
"cost_boundary": { "max_spend_per_action": 200, "requires_human_approval": true },
"loop_boundary": { "max_iterations": 6, "has_escape_hatch": true }
}Pick a Trust Level — the manifest updates live.
Why governance matters
A prompt can ask. Only a gate can stop.
Prompt · advisory“ask a human before issuing a refund”
no barrier
crossed the threshold → refund executed
Gate · enforcedrefund tool requires an approval token
approval gate
held at the gate → escalated to a human
The model can be jailbroken, confused, or wrong. The boundary lives outside the model, so it holds anyway — that’s what an AgentAz Trust Level encodes.
Free tools
All toolsBrowser-based, deterministic, no login — run them on your own agents, not just ours.
Compliance ScannerGrade any agent's governance in ten seconds.Agent WatchTrack your agent's Trust Level and catch drift.Tier BadgeGenerate a Trust-Level badge for your repo.Risk AssessmentClassify your agent's autonomy and blast radius.Predict the BreakA game that sharpens your eye for agent failures.
Featured Defensive Kits
Browse allFlagship blueprints with documented worst-case actions, failure modes, and Trust Levels — the AgentAz™ reference set.
A sales-research agent that builds a cited account brief — firmographics, tech stack, triggers, key people — flags stale data, and never fabricates contacts or triggers.
ClaudeLangGraphOpenAIIntermediate
–Trending
An action-item agent that extracts owners, tasks, and due dates from meetings, tracks status, and flags ambiguous ownership instead of inventing it.
ClaudeLangGraphOpenAIIntermediate
–
Flagship
AI Bug-Fix & Draft-PR Agent
An AI bug-fix agent: reproduces an issue, finds root cause, writes a minimal fix plus tests, and opens a focused draft PR — sandboxed, small-diff, human-gated.
ClaudeLangGraphOpenAIAdvanced
–
Flagship
AI Contract Review Agent
A contract-review agent: reads clause by clause against your playbook, flags risky and missing terms with severity, and proposes fallback language. Not legal advice.
ClaudeLangGraphOpenAIAdvanced
–
An AI support agent that grounds in your KB and CRM, then resolves, drafts, routes, or escalates — with confidence thresholds, refund caps, and human review.
ClaudeLangGraphOpenAIAdvanced
–Trending
Flagship
AI Incident Response Agent
An on-call SRE agent: correlates alerts with metrics, logs, and recent deploys, proposes safe mitigations, drafts status updates, escalates SEV1 — with approval gates.
ClaudeLangGraphOpenAIAdvanced
–
Flagship
AI SOC Alert Triage Agent
A SOC tier-1 agent that enriches and correlates alerts, scores severity with MITRE mapping, and recommends contain, dismiss, or escalate — never auto-containing blindly.
ClaudeLangGraphOpenAIAdvanced
–Trending
An alert-tuning agent that scores alerts by actionability, dedupes, and time-box-suppresses proven noise — never silencing an alert tied to a real incident.
ClaudeLangGraphOpenAIAdvanced
–Trending
A marketing agent that builds a structured campaign brief from your inputs, flags assumptions and missing data, and never fabricates audience stats or guarantees results.
ClaudeLangGraphOpenAIIntermediate
–
A community triage agent that routes and answers incoming questions from your knowledge base, flags duplicates, and sends spam, abuse, and safety issues to moderators.
ClaudeLangGraphOpenAIAdvanced
–
A policy Q&A agent that answers employee questions only from the official handbook with citations, flags what isn't covered, and routes sensitive HR matters to a human.
ClaudeLangGraphOpenAIIntermediate
–Trending
Browse by category
All blueprints30 categories across 5 domains — pick where your agent lives.
Customer-facing
Engineering & Data
Operations
See the loop run — with the gate enforced
Watch a blueprint’s control loop reason, call tools, and stop at the human-approval gate before any irreversible action. This is the same enforced loop that ships in every runnable kit.
Incident Response Agentlive preview · mock tools
inputALERT checkout-service 5xx 0.4%→9% over 6m · deploy v812 finished 4m earlier
Latest additions
Maps each change to your control catalog, assembles tamper-evident evidence, and gates filing behind a human — it records, it doesn't decide compliance.
ClaudeLangGraphOpenAIAdvanced
–Trending
Pulls logs, metrics, traces, and recent changes into one timeline, ranks root-cause hypotheses, and gates any published RCA behind a human. Diagnoses, never remediates.
ClaudeLangGraphOpenAIAdvanced
–Trending
An access-request agent that checks role policy and separation-of-duties, auto-provisions low-risk access, and escalates privileged or sensitive requests.
ClaudeLangGraphOpenAIAdvanced
–
A sales-research agent that builds a cited account brief — firmographics, tech stack, triggers, key people — flags stale data, and never fabricates contacts or triggers.
ClaudeLangGraphOpenAIIntermediate
–Trending
An action-item agent that extracts owners, tasks, and due dates from meetings, tracks status, and flags ambiguous ownership instead of inventing it.
ClaudeLangGraphOpenAIIntermediate
–
Flagship
AI Bug-Fix & Draft-PR Agent
An AI bug-fix agent: reproduces an issue, finds root cause, writes a minimal fix plus tests, and opens a focused draft PR — sandboxed, small-diff, human-gated.
ClaudeLangGraphOpenAIAdvanced
–
Weekly Highlight
AI Incident Response Agent
An on-call SRE agent: correlates alerts with metrics, logs, and recent deploys, proposes safe mitigations, drafts status updates, escalates SEV1 — with approval gates.
- Trust Level
- A4 — Limited Autonomy
- Worst-case action
- Executes an allowlisted low-risk, reversible remediation step (e.g. restarting a stuck service) that turns out to be unnecessary. Every auto-executed step is sandboxed with a registered rollback. Irreversible or high-impact actions — production rollbacks, scaling, security or config changes — are never auto-executed; they require human approval.
- Guards against
- Misdiagnoses the incident and targets the wrong service with a remediation step.