agentkit — open-source AI agent toolkit

// context & bottleneck

Frameworks make the demo easy; production makes the demo lie. The model sometimes returns prose where you asked for JSON, a runaway tool loop is a bill, and two tenants must never see each other's documents. These aren't exotic failures — they're what broke while shipping real agents for paying clients.

// solution & architecture

Extract the reusable core from two production systems — a multi-tenant dental copilot and a social publishing bot — into one zero-dependency library, and leave the client-specific parts private.

›An agent loop with a hard budget: token and cost ceilings are checked before every model call, so a runaway agent stops before it spends — verified by tests that drive the loop into the cap on purpose.

›Reliable structured output: ask for JSON, validate the shape, and on a miss re-ask with the exact error so the model self-corrects instead of the pipeline crashing.

›Tenant-safe RAG: retrieval is scoped by a tenant key at the store level, not filtered after the fact — a query can never see another tenant's chunks.

›Typed tool registry that never trusts model shapes: garbage arguments coerce to {}, and a tool refuses cleanly instead of throwing.

// business impact

Released open source under MIT, tested with node:test against a mocked LLM and verified end to end against a live model — structured extraction, a tool-calling loop, and budget accounting, with the reasoning-token trap that emptied replies in production diagnosed and fixed in the open. This is the reusable, tested part of what makes the client systems work.