agents · workflow

Make every repair earn stronger evidence

Intent becomes a candidate; deterministic feedback narrows the problem; vendor tooling is used only when the target environment itself needs to establish the result.

01

Goal

Engineer defines the behavior.

Human intent
02

Propose

Agent produces a candidate change.

Model output
03

Fast check

2ot returns deterministic diagnostics.

Compiler evidence
04

Context

Source and semantics bound the repair.

Program evidence
05

Repair

Agent revises from observed evidence.

Model output
06

Vendor check

Target tooling establishes acceptance.

Vendor evidence
illustrative boundary bug

Compilation can succeed while the requirement is still wrong

// candidate
SugarFeeder := SugarLevel <= 10.0;

// requirement
SugarLevel < 10.0   => TRUE
SugarLevel >= 10.0  => FALSE
Candidate exists
No independent verification yet.
INFERRED
2ot compiler accepts it
Syntax/type evidence, not requirement proof.
ANALYZED
Target toolchain accepts it
Engineering-environment evidence.
VENDOR
Machine behaves correctly
Requires runtime or physical observation.
SEPARATE
INNER LOOP

Keep repeated checks cheap

2ot check --fast is the tight repair loop before vendor startup becomes necessary.

ESCALATION

Use the strongest authority you actually reached

Compiler, vendor, controller and physical evidence remain distinct claims.

Diagnostics →Agents & MCPSafety model