Skip to Content
Autonomy & gates
last verified · 2026-07-16

Autonomy & gates

Say you ask Lumis to fix a typo on a settings page. An agent doesn’t stop and ask you to approve the fix, then ask again to approve running the tests, then ask again to approve opening a review. It plans the change, makes it, runs the tests, gets it reviewed, and moves it forward — all without you in the loop. But if that same session hits a change that would alter your production database’s schema, it stops. However confident the agent is, it waits for you.

That’s autonomy and gates working together. An agent earns the right to run a chain of work on its own by clearing a bar on every single cycle of that work — and a short list of actions never clears the bar, no matter the score.

The confidence gate, mechanically

Before an agent proceeds to the next cycle of a chain it’s running unsupervised, that cycle is scored against five factors. Each factor gets a score out of 20, for a total out of 100, and the score is written down — every factor, every time, not just the total.

FactorThe question it answers
GroundingIs there solid context behind this, or is the agent guessing?
PrecedentHas work like this been done successfully here before?
ReversibilityIf this is wrong, can it be undone cheaply?
ValidationIs there a real check — a test, a build, a diff review — that would catch a mistake?
RiskIf this is wrong, how much does it affect — one file, or everything?

These are the same five questions a careful person asks themselves before acting without checking in first. Lumis just asks them explicitly, on the record, on every cycle — instead of trusting an agent to have asked them silently.

The published bar

An agent may proceed to the next cycle on its own when the total score is 80 or higher out of 100. Below that, the chain pauses and surfaces what’s missing — which factor was weak, and why — so you can supply what the agent couldn’t find on its own.

That’s the whole rule for reversible, in-scope work: score 80 or above, proceed and log it; score below 80, stop and ask. No hidden exceptions, no per-agent tuning that isn’t visible to you.

The hard gates — no score buys past these

A handful of actions never proceed on a confidence score, full stop. These are the things where being wrong is expensive or permanent enough that a number isn’t the right kind of answer:

  • Deploying to production.
  • Spending over a limit you’ve set.
  • Touching secrets, API keys, or credentials.
  • Any destructive or irreversible operation.
  • Reversing a decision that’s already been locked in.

An agent can prepare, stage, and recommend any of these. It cannot execute one. Only you can.

Worked example: two tickets, two outcomes

Ticket: Fix a typo in the settings page description.

Grounding

The exact wrong word is visible in the current page; no ambiguity about what to change.

Precedent

Text-only edits to this page have shipped before with no incident.

Reversibility

A one-line text change on a branch; trivial to revert.

Validation

The page’s existing tests re-run against the change and pass.

Risk

Affects one string on one page. Nothing else depends on it.

Score: well above 80. Result: the agent makes the change, runs the tests, gets it reviewed, and moves it to Done — no pause.

The difference isn’t how confident the agent feels. It’s whether the action is the kind a number should ever be allowed to authorize.