The harness: a loop that survives
In short Keep the model's loop, change where it runs: durable, one active turn per conversation, interrupts as signals. Let the model steer and delete the routers, lanes and planners that tried to help.Full verdict ↓
The harness is everything between "a message arrives" and "the reply is delivered": the loop that calls the model, runs its tools, folds in new input, stops, and survives a crash on the way. I did not invent the loop. I started from an existing open-source agent and changed where it runs.
The design
The journey
- Extend, don't replace. Keep an existing agent's loop and tools; fix where it runs. Its loop ran inside the chat process and kept state in one local file.
- Leverage, don't rebuild. Stop writing file and web tools. Delegate coding to coding agents and orchestrate them.
- Delete the helpers. Classifier-driven lanes, routing and a planner were removed. The model decides how to handle a request.
- Delete the modes. A conversational skill and session-mode idea was removed outright.
- One kind of delegated work. Tasks replaced subagents.
- An agent chooses its harness. The platform can host a coding agent for an agent and translate its events, instead of copying it.
Decisions that held
| Decision | What changed | Why |
|---|---|---|
| Keep the loop, move it out of the chat process | The loop became a durable workflow; state moved from one local file to a shared database | One process held the only copy of the conversation guard, and one file serialized every write |
| One active turn per conversation | A long-lived guard per conversation holds only a pointer to the running turn | A new message is forwarded into the running turn, or starts one, with no in-memory lock |
| Interrupt by cancelling, not abandoning | In-flight work is asked to stop and clean up | Abandoned work keeps running unseen. The cost is honest: interrupt latency is bounded, not instant |
| The model steers, code only guards | A handful of rails: approvals, spend, context size, recall | A router or planner decides worse than the model reading the request |
| A failed tool is an observation | The error goes back to the model | The model can reason about a failure like any other result |
| Tasks, not subagents | Delegated work has a durable record, can be cancelled and reports back | "Needs this result first" is a dependency between tasks, not an inline wait |
| Host a harness, don't copy it | A coding agent runs in a sandbox and its events are translated | A dedicated coding agent is better at coding than anything I would write |
What already exists
| Option | What it gives | What it leaves |
|---|---|---|
| Agent frameworks (LangGraph, OpenAI Agents SDK) | A way to express the loop and its branches. LangGraph checkpoints between steps | Crash recovery is lighter than full replay, or comes from a runtime you add |
| Durable runtimes (Temporal, Restate, Inngest, DBOS) | Resume after a crash, retries, long waits | They run work, they do not know what an agent is |
| The agent I started from (Hermes Agent) | A good loop and tools, gateways to many chat platforms | Its loop runs in the chat process and its state is one SQLite file |
| Mine | A durable loop with a conversation guard, interrupts and tasks | A thin layer I maintain myself |
Most comparisons I found are written by the vendors, so I leaned on the design problem and not on their rankings: single-service agents are often fine on a framework's checkpointer, and a durable runtime pays for itself when you need long waits, interruption and recovery across services.
Where it stands, and my verdict
| Component | Status | |
|---|---|---|
| Memory (post 2) | Built, traceable; two memory journeys pass | thin |
| Tools (post 3) | Running; journey suite measures discovery, chaining and failure | measured |
| Loop, durability, interrupts | Running through every later change | stable |
| Failure handling | A tool that fails once is retried; one that stays down is disclosed | checked |
| Router, lanes, planner | Removed. A classifier now only suggests a label the model may ignore | deleted |
| Hosting a coding agent | Designed and partly built; the platform only translates | new |
Verdict: the harness got smaller every time I touched it. Each rewrite moved intelligence out of the harness and into the model, and kept the harness for the things that must be guaranteed: not losing work, not running two turns at once, not spending without limit. The journey suite now exercises the failure path: a flaky tool is retried, and a tool that stays down is disclosed rather than invented. Next: what the model actually reads on each turn.