Agentic Execution Engine
I designed an execution engine that drives multi-step plans to completion without a workflow server. The agent follows the controller pattern — read desired state, read actual state, close the gap, terminate. Stateless, crash-recoverable, and plans stay editable at any time.
The Problem
Real-world agent tasks involve 5-50 steps across days or weeks, mixing automated and manual work. Plans change during execution — users discover new requirements, add or remove steps. Traditional workflow engines treat this as an exception, not the norm.
The Insight
Think of a thermostat: it reads the target temperature and the thermometer, turns the heater on or off, and forgets. If it reboots, it reads both values and continues perfectly. The orchestration agent works the same way — stateless reconciliation on every invocation.
Key Design
Separate command from sync. The execute tool creates jobs but doesn't write the plan. The sync tool reconciles all state in one write. This makes parallel execution safe — no concurrency conflicts.
Why Not a Workflow Engine?
Temporal, Airflow, and Step Functions all require the workflow definition in code. Our plans are human-readable markdown — editable by anyone, anytime, with no code deployment or server restart needed.
Interactive Demo
Stateless Reconciliation Engine
Watch the controller pattern drive a plan to completion
Checkpoint/Replay (Temporal, Step Functions)
Stateless Reconciliation (This Design)
The Reconciliation Loop
Read
Download the plan (desired state) and query all jobs (actual state)
Diff
Compare desired vs actual. Find steps that are unblocked and ready to execute
Act
Create jobs for unblocked steps, sync statuses back to the plan in one write
Terminate
Agent exits. No persistent process. Next event triggers a fresh invocation