conceptJun 20, 2026
Agentic Workflow Engine
A framework for building bounded, observable, human-in-the-loop agents — budgets, tool gating, and full execution traces by default.
AgentsLangGraphObservabilityPython
The reliability patterns from Designing Reliable AI Agents, turned into a reusable engine.
The idea
Most agent frameworks optimize for the demo — how quickly you can wire a model to tools. This one optimizes for the operator: what happens on step 40, when the budget runs low and the agent wants to do something irreversible.
Core primitives
- Step & token budgets enforced at the loop level, with graceful exhaustion.
- Tool classification — reversible tools run freely; irreversible ones pass a confirmation gate.
- Structured traces for every decision, replayable after the fact.
Status
Design phase. We are validating the primitives against real client workloads before opening it up. Following along? The blog is where the thinking gets published first.