craft-a-bot
An interactive, browser-based agent simulator: assemble modular bricks — brain, sensors, memory, tools, actuators, safety — on a virtual workbench to build agents that pursue goals in a simulated environment, with a six-chapter tutorial covering observation, memory, tool use, hallucination, and human oversight.
Last reviewed: 2026-08-24
What it is
A monorepo (SvelteKit, TypeScript, Turborepo) that treats AI agent construction as a physical toy: pick a brain cartridge, wire in memory and tools, set a goal card, and watch the agent act — with every prompt, decision, and tool invocation visible in a real-time flight recorder. The default Demo Brain is scripted and needs no API key; an OpenAI provider is available for anyone who wants to bring their own key.
Why it's here
Most agent tutorials teach the happy path. This one is built around the failure modes this site's Security pillar and risk library already name — hallucination, excessive agency, missing oversight — as first-class parts of the toy, not a compliance afterthought bolted on at the end.
Governance notes.
What this demonstrates about governed AI — not a feature list, a connection back to the same discipline the rest of this site argues for.
Tamper-evident execution traces
Every run produces a complete, exportable trace with SHA-256 digests, so what an agent actually did can be verified after the fact rather than taken on trust.
Capability scoping and step budgets
An agent is bounded in what it can do and how many steps it can take before it stops — the concrete mechanic behind "bound what actions are possible before relying on what actions are likely."
Optional approval mode
A human-in-the-loop checkpoint before higher-consequence actions proceed — the same control pattern this site recommends for any agentic tool with real-world effect.
Public code, not a live demo — yet.
No hosted demo is embedded here. The repository is the thing to read.