Agentic AI platform
An agent system with adult supervision: planner and workers, a cross-model critic, guard-railed tool calls, human checkpoints for irreversible actions and deterministic replay.
What this design does
Agents fail in ways chatbots can't: they act. This platform wraps the acting in discipline — an orchestrator owns the budget and delegates to workers; a critic model from a different vendor reviews drafts (models share their own blind spots, so the reviewer shouldn't share the author's); every tool call passes schema guardrails and an allow-list; irreversible actions pause for a human; failed multi-step actions are compensated saga-style; and every step lands in a run log that makes any trajectory replayable, deterministically, for debugging.
The patterns it composes
- Orchestrator-worker — one agent owns the goal and budget; workers own steps — no anarchy
- Planner-executor — plan first, then execute steps — re-plan when reality disagrees
- Reflection / critic loop — a DIFFERENT vendor's model reviews output; one revision round, hard cap
- Guardrail / tool-router — every tool call is schema-validated against an allow-list before it runs
- Human-in-the-loop checkpoint — irreversible actions (money, deletion, email) require a human click
- Saga compensation for tool actions — a failed step 3 of 5 triggers compensating calls for steps 1-2
- Deterministic replay of runs — every step is logged; any run replays exactly for debugging and evals
Import it, attach your own numbers, and verify the design against failures — or read the engineering notes.