Case study · located work
ProgressGate
An npm package that detects semantic stagnation in agent tool-calling loops, published with tests, a quickstart, and a live demo build.
ObservedAshutoshVJTI
- Problem
- An npm package that detects semantic stagnation in agent tool-calling loops, published with tests, a quickstart, and a live demo build.
- Jev’s role
- One Jev systemOne call per check returns six atomic signals (assumption contradicted, strategy novelty, material progress, and others); Jev never executes or blocks a tool call itself.
- Architecture
- Agent trajectory → Jev semantic signals → deterministic hysteresis policy in application code → CONTINUE/WARN/REPLAN/HALT decision.
- Why it matters
- Keeping the halt decision in a deterministic policy layer, with a fail-open default on Jev/network errors, means an API outage cannot itself look like agent stagnation.
Code boundary
Jev answers the bounded questions described above. Thresholds, retries, side effects, and anything that must be reproducible stay in the surrounding application code. The research record does not capture this project’s exact question set or threshold values; read the linked source before copying the architecture.
Sources
Limitations
- A located repository or demonstration does not establish production reliability or independent validation.
Reusable pattern and related records
- Confidence gate PlausibleAutomation proceeds above a threshold; ambiguous cases go to another check or a person.
- Agent Loop Stopper Authored HypothesisA loop controller deciding continue, retry differently, verify, ask, or stop after every step.
- Cheap decision calls could make tens or hundreds of semantic judgments per event economical. SpeculativeThe product design follows from published pricing and parallel question primitives, not independent production evidence.
- Probabilistic predicate + deterministic action PlausibleJev supplies fuzzy predicates while TypeScript, policies, and workflows execute constrained actions.