Deterministic reflex runtime
Nothing fast stops to think.
Animals develop reflexes to survive. Athletes develop them to win. Reflexes exist because reasoning is too slow to repeat. Cerebellum records the path your agent worked out once, collapses it into a reflex, and replays it without the reasoning model.
First run · your model drives
32.0 s·6 model calls, done
120k tokens in·1.8k out·$0.25
Run 100 · Cerebellum drives
≈ 650 ms·same 5 tool calls, 30 ms round trip before each
0 reasoning tokens·✓ ×16 runs so far, one fired every 2 s
Reflexes start guided, handed to your model as a plan, and graduate to executed once they have proved themselves. The run Cerebellum drives is an executed reflex, which is where the product is heading; guided reflexes come first, and save tokens while the model still drives every step.
Stimulus → local reflex → propagation → collective response
No fish sees the whole school. Each one reacts to its neighbours, and the response crosses the group faster than any of them could stop to think. Switch the reflexes off to see what thinking it over costs.
The film · 1:30
One task, followed until it becomes a reflex.
Run 1 of a task, chasing overdue invoice INV-1842: your model drives six turns of reasoning around five tool calls, get_invoice, get_customer, check_payments, draft_email and send_email, and takes 32.0 seconds, 120,000 tokens in and $0.25. Run 100, an executed reflex, makes the same five calls with a 30 millisecond round trip before each and no reasoning tokens, in about 0.65 seconds: about 50 times faster on this example. 01, observe: every run is recorded, the failed ones too; of 100 runs, 87 took the same five calls, 7 took other paths and 6 failed and were kept. 02, collapse: what repeats is kept and the reasoning drops out, leaving the five tool calls. 03, approve: a person approves the reflex and every version is immutable; changing one argument of draft_email makes a version 2 that waits for its own approval. 04, earn: a reflex starts guided, handed to your model as a plan, while its adherence and success rate are measured, and once both clear 95% it is promoted to executed, where no reasoning model runs between the calls. Guided comes first; executed is where the product is heading. 05, gate: every tool call is checked against policy first. The three reads are allowed, draft_email is modified to copy in accounts@example.com, and send_email, for an invoice over £5,000, asks a human and is approved by finance. 06, hand back: when check_payments returns an unexpected partial payment, the reflex stops at step 3 and hands the task back to your model with steps 1 to 3 attached, so nothing runs twice. Your harness keeps its own model and tools, and Cerebellum is called at three points: before the task, before each tool call and after the task. It ends on the line: Nothing fast stops to think.
How a reflex forms
Nobody writes the reflex. The work does — and you approve it.
- 01 / OBSERVE
Every run is recorded
The tools called, the order, the arguments, the outcome. A hundred runs of the same task leave a hundred traces, and the failed ones are kept too: they are what stops a bad routine becoming a reflex.
- 02 / COLLAPSE
The reasoning steps drop out
What repeats every time is kept. What only existed to decide the next step is discarded.
- 03 / REPLAY
One fixed path, run as code
Same five tool calls, no reasoning between them. When something unexpected appears, the model is called back in.
More than a lookup
What Cerebellum does that a classifier can't.
- 01
Finds the routine for you
Transcripts from every run are normalised and compared, and repeated tool sequences surface as candidates. Nobody has to notice the pattern.
- 02
A person approves it
Nothing becomes a reflex without a developer's sign-off, and every version is immutable once approved. A change is a new version and a new approval.
- 03
Earns its autonomy
A reflex starts guided, handed to your model as a plan, while its adherence and success rate are measured. Only one that proves itself is promoted to executed, where no reasoning model is involved.
- 04
Every tool call is gated
Policy is checked before any side effect, including each step of a reflex: allow, deny, modify, ask a human, escalate. Every decision is recorded with its reason.
- 05
Knows when to stop
Unexpected output, low confidence or a changed policy hands the task back to your model with the completed steps attached, so nothing runs twice.
- 06
Stays out of the model path
Your harness keeps talking to its own model and tools with its own credentials. Cerebellum is called at three points: before the task, before each tool call, and after the task.
Two paths, one stimulus
A brain doesn't think about everything.
- Reasoning
- The signal climbs into the cortex and searches for a path, one step at a time. It gets there, but slowly, and by a different route each time.
- Learned reflex
- Repetition wears a path. Once the cerebellum has learned the response, the same signal turns there and goes straight back down: no search, no pause, the same path every time.
Cerebellum in figures
- faster on the five-step example above
- ≈ 50×
- a day of model spend it removes, at 1,000 runs of that example
- ≈ $250
- reasoning tokens on the known path
- zero
- tool call checked against policy first
- every
Illustrative, not a benchmark: the speed-up and the saving are the worked example at the top of this page.
Why give your agents reflexes
Repeated work stops costing what novel work costs
Reasoning is priced per token and per second. A reflex is priced like code.
The same input stops producing a different plan
A reflex is a fixed path. You can test it, audit it and trust it to behave the way it did yesterday.
Your models go back to doing the hard part
When routine work runs itself, frontier reasoning is spent on the exceptions that actually need it.
Security
Built to sit in front of your tools.
Your credentials stay in your harness
Tools and model steps run where they always have, with your keys. Cerebellum decides what happens next; it never holds the means to do it.
Strict tenant isolation
Every query, cache key and workflow is scoped to your organisation and project, taken from the authenticated context and never from the request.
An audit trail that cannot be edited
Approvals, policy decisions and key changes are written to an append-only log, enforced in the database itself and not only in application code.
Give your agents a cerebellum.
Keep frontier models for the work that genuinely requires frontier reasoning.