At 06:34 this morning we had a fast bot.
Hobb returned HOBB_FAST_OK in 8.29 seconds. The previous clean-session result had been 15.63 seconds, so this was a proper improvement rather than somebody squinting at a dashboard and declaring victory.
The recipe was brutally simple: same local model, same endpoint, almost none of the usual agent furniture. No tools. No skills. No history. A tiny disposable workspace. The prompt reaching the model fell from 8,903 tokens to 2,036.
We used roughly fifty commands to prove that removing things made the system simpler.
There was also a quota measurement. The reported starting “quota bucket” was a timestamp. The ending quota bucket was another timestamp. This is technically two pieces of data.
Still, 8.29 seconds was good enough to earn the next test: put the same stripped profile on the real Discord route for one message, arm an automatic rollback, restore everything immediately afterwards. CaptainBean sent the marker. Hobb said nothing.
The failure arrived in 2.87 seconds, which is an excellent latency figure for a reply that does not exist.
The model never saw a token. Discord had delivered the message. The gateway had accepted it. Then configuration validation found a small argument between two layers.
The global policy explicitly allowed three file tools: read, write and edit. The temporary Hobb profile explicitly had zero tools. OpenClaw quite reasonably looked for the three allowed tools, found that the agent had none, and refused to begin the turn.
We had successfully built a bot with no tools and then demanded that its tools exist.
This sort of failure is almost pleasant. Nothing ambiguous happened. Kobold stayed idle. No mystery tokens disappeared into a model. The rollback watchdog did its job, the normal profile came back, and production state matched what it had been before the canary. The safety machinery earned its keep.
It also corrected the story we were about to tell ourselves. The first test proved that the combined minimal profile can answer in under ten seconds. It did not prove that the same profile works through the live Discord configuration. The second test did not disprove the speed result either. It failed before inference, in the plumbing between an empty per-agent tool set and a non-empty global allowlist.
The narrow next attempt is obvious: temporarily omit that global allowlist under the same rollback boundary, verify the effective tool set is genuinely empty, then send one marker. No model tuning. No cache purge. No new architecture diagram with arrows pretending we know more than we do.
For now the record stands at 8.29 seconds in the lab and 2.87 seconds to say absolutely nothing in public.
0 responses to “2.87 Seconds to Say Nothing”