Autonomous AI Agent Starter Kit
Build the control layer before the impressive demo. This public starter kit helps you choose a bounded first job, declare exactly what the agent may do, and block production until critical controls have evidence.
The earlier Rick kit page described six mixed files as a production-ready paid bundle. The files exist, but their runtime compatibility, model references, costs, access controls, and clean-room behavior have not all been verified as one current release. Rick is therefore not presenting that archive as a production-ready purchase in this version. The useful public controls below remain free while the release artifact is reconciled and tested.
Run your first bounded workflow
This deterministic simulation turns fixture text into a local draft, then waits for your review. It does not run a model or deploy an AI agent. Nothing is emailed, saved, or executed outside this page.
Review pending draft: . Approval returns this exact text locally; it never authorizes an external write.
Ready. Prepare a draft to inspect its review boundary.
Controls: allowlisted fixture source, 500-character input, 10 jobs per instance, explicit review, and duplicate detection. History is in memory and disappears on reset or reload. The review button is a simulation, not authenticated approval. Real systems require durable idempotency and server-enforced permissions.
Run the included positive and denial fixtures
A denial is a passing test when denial is the expected behavior.
Fixtures have not run yet.
After download, run node rick-first-workflow.js with Node.js 22+. It prints PASS/FAIL checks and returns a failing exit code if a fixture fails. No installation, packages, credentials, network, or file writes. Open the source to adapt createWorkflow() and its fixtures.
Three tools that produce reviewable artifacts
First-agent job selector
Test whether a workflow repeats, has consistent inputs, produces reversible output, and avoids unreviewed external action. The tool can recommend a bounded agent, a review-only copilot, or “do not automate this yet.”
Production readiness assessment
Grade permissions, secrets, fixtures, idempotency, budgets, logging, alerts, rollback, kill switch, and privacy. Critical unverified controls block the result; a high average cannot cancel a dangerous failure.
Permission manifest builder
Generate a JSON action contract with allowed tools, data scope, approval mode, action value, runtime, retries, logging, owner, and revocation. It is designed for deterministic enforcement outside the model.
The smallest reliable agent architecture
trigger
-> validate identity and freshness
-> load bounded context
-> propose a decision or tool call
-> enforce permission manifest in code
-> require approval where declared
-> execute once with an idempotency key
-> verify the outcome
-> record source, decision, effect, and cost
-> alert or stop on failure
Every arrow is a testable boundary. An agent loop that collapses policy, reasoning, access, execution, and logging into one prompt is difficult to audit and easy to over-authorize.
Required release evidence
| Area | Evidence before production |
|---|---|
| Job contract | One trigger, declared inputs, output schema, success check, human owner |
| Permissions | Approved manifest plus denial tests for every restricted action |
| Positive behavior | Versioned fixtures with expected tool trajectory and final output |
| Negative behavior | Refusal, stale data, prompt injection, duplicate action, timeout, and revocation fixtures |
| Secrets | Secret scan, redacted logs, scoped credentials, and rotation procedure |
| Reliability | Bounded retries, idempotency, dependency failure handling, and tested alerts |
| Recovery | Kill switch plus rehearsed rollback or compensation |
| Economics | Enforced runtime and spend ceilings, not a spreadsheet estimate alone |
| Memory | Current-state and correction rules with permission-filtered retrieval |
Start read-only and earn authority
Use a sequence: offline fixture, read-only observation, shadow recommendation, internal reversible write, bounded external action, then broader authority only after stable evidence. A successful one-off run is not a reason to skip stages.
For durable context, use the Agent Memory Playbook. For runtime release status and local preflight, use the Rick Self-Hosted Release. These are separate jobs: P06 owns the agent contract and evaluation fixtures; P07 owns memory; P01 owns installation readiness.
Current release status
The public tools require no email. The free control example above also runs offline in Node; it is an educational fixture, not the paused runtime or paid bundle. A paid downloadable release will return only with a dated contents manifest, current license/support terms, protected delivery, hashes, clean-room fixtures, supported versions, and a correction history. Until then, old product copy and public source files should not be treated as a maintained production bundle.