What I deleted
In short I deleted nearly as much as I kept, and the reasons repeat: the model did it better, it duplicated something, it sat in the wrong layer, or it was heavier than the job. The design is what survived.Full verdict ↓
Every post so far described something that exists. This one is about what does not. Going back through my notes I found thirteen things I built and later removed. The pattern in why is the real design.
The record
Why things went
| Reason | What went | The lesson |
|---|---|---|
| The model did it better (3) | Classifier-driven lanes, routing and a planner. A conversational session mode. A category filter on skill search | If the model can read the request, a hand-written helper in front of it only adds a way to be wrong. I delete the helper, not the model |
| It duplicated something (4) | Intentions for waking. Subagents next to tasks. A message kind column next to role and speaker. Wake and watch tools in the harness |
Two ways to do one thing means one is stale. Find the existing mechanism and use it |
| Wrong layer (2) | A second door on the toolbox for code. Sandbox types chosen at creation | Put a decision where its information lives. Privilege belongs with the credential, not with a kind of box |
| Heavier than the job (1) | A microVM per session | A pure function should not pay for a kernel |
| Never earned its keep (3) | The mining and graph memory pipeline. Persona and trait lenses. A suite for supervisors answering questions | If I cannot show it helps, it goes. That is the whole reason evals exist |
The rules that were left standing
| Rule | Held in |
|---|---|
| The model steers, the harness supports | Harness, waking |
| Leverage, don't rebuild | Harness, toolbox |
| Everything durable | Harness, memory |
| Every belief cites its source | Memory |
| The agent never holds the keys | Toolbox, sandbox |
| Measure, then delete | Evals, this post |
Where it stands, and my verdict
| Component | Status | What is open |
|---|---|---|
| Memory | Built, traceable; two memory journeys pass | No memory benchmark. Do conclusions earn their cost? |
| Tools | Running; journey suite measures discovery, chaining and failure | Code over tools is unevaluated on real requests |
| Harness | Running, smaller than ever | Hosting a coding agent is designed, partly built |
| Context and skills | Running | The skill gain is expected, not re-measured |
| Runtime and sandbox | One day old | A type check before a program runs |
| Waking and asking | Small and live | Rules that stop a run |
| Evals | Four suites; journeys are the main test. 35B 72/81, 8B 32/81 | A memory benchmark, a code-mode eval, and the production prompt on every model |
Verdict: an agent that runs on its own is mostly a set of things it is not allowed to do, and a set of things I stopped building. The loop, memory and toolbox kept their shape. Almost everything else was rebuilt or removed, usually within weeks and sometimes within a day, and each deletion made the system easier to reason about. The weakest part was the one that would have saved the most work: measuring earlier. The first eval suite was itself a deletion candidate: its 9 of 24 ceiling was a measurement fault, not a model limit, and fixing it changed the model verdict. If I did this again I would write the evaluation before the component.
What I will do next, in order: evaluate memory and code mode, because neither has a dedicated test and two memory journeys are a signal, not a benchmark. Run the production prompt on every model, because the suite's own prompt is not the one the agent uses; Gyro-S fell from 71 to 62 on it, and the 8B production run is missing. Make a rule stop a run, because today I can only watch the spend.