Build blueprint
Agent Tool Firewall
A local gateway that scores every proposed tool call and requires confirmation only when risk is high.
- Problem
- Autonomous agents can select destructive or irrelevant tools from large catalogs.
- Why Jev
- Tool execution needs a typed allow/deny/escalate decision with confidence, not generated prose.
- Architecture
- Agent proposal → Jev risk and intent questions → deterministic policy thresholds → tool or approval UI.
- Current alternative
- Prompt-based guardrails, static allowlists, or an LLM judge on every call.
- Jev advantage
- A decision can run inside every agent step while preserving explicit policy code.
- Unknowns
- Calibration under prompt injection and whether confidence remains stable across tool schemas.
Why this confidence: The typed-decision interface directly matches tool gating, though safety still needs deterministic backstops.
This is an authored hypothesis derived from the research corpus. Nothing here demonstrates product demand, or that Jev performs well on this particular workload. Run the validation experiment before building past the MVP.
Sources
Limitations
- Calibration under prompt injection and whether confidence remains stable across tool schemas.
- This is a research hypothesis, not evidence of product demand or Jev performance in this workflow.
Supporting research
- Confidence gate PlausibleAutomation proceeds above a threshold; ambiguous cases go to another check or a person.
- typesafe-mcp ObservedAn MCP server exposing TypeSafe evaluation to coding agents.
- 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.