Back to Home
Project 06

AI Agents

AI that takes real action, safely

A framework for AI agents that carry out real tasks through controlled tools, safety checks, and a full record of what happened. It keeps suggesting separate from doing, and handles retries safely.

Role
Systems designer
Status
In active development
Period
2024 — Now
Domain
AI infrastructure
LLM OrchestrationTool-callingPydantic v2StreamingIdempotency
The Problem

Most AI agents blur the line between suggesting and doing — they either just chat, or they quietly take actions with real consequences. What's missing is a disciplined setup where every action can be reviewed before it happens.

Approach

Keep thinking and doing separate. Each tool declares up front what it will do; the system stages the action and waits for a human or a policy to approve it before running. Retrying is safe by design — nothing gets done twice by accident.

Architecture

Tool contract

Pydantic v2 schemas describe inputs, side effects, and rollback semantics for every tool.

Runtime

Streaming LLM orchestration with deterministic tool dispatch, structured logs, and a replayable audit trail.

Guardrails

Policy hooks veto unsafe actions before they reach the proposer; idempotency keys make retries safe.

Extensibility

30+ tools shipped; a new tool lands via a single class plus a JSON spec — no runtime rewiring.

Outcomes
  • Every action is reviewed before it runs — nothing happens silently.
  • New capabilities plug in fast, without rewiring the system.
  • Safe to retry — the same task never runs twice by accident.
Stack
LLM OrchestrationTool-callingPydantic v2StreamingIdempotency keysAudit trail
Screenshots
Runtime flow — propose, guardrail check, approve, then execute.