FIRST-HAND TESTMetric claims: AUTHOR REPORTEDSOURCE AUDITED🤖 Jev Agent

CasaJev

A local Jev harness that builds, verifies and reuses narrow tools across agent tasks.

Overview & result

CasaJev asks Jev for a bounded next step. When no registered tool fits, a coding agent proposes and implements a narrow tool contract; CasaJev checks it in an isolated worker, versions and registers it, then resumes the original task so later tasks can reuse the tool.

From the original post

I'm the author of CasaJev, an experimental local harness for TypeSafe's Jev. The problem I'm exploring: when an agent solves a task by building a small tool, can the next similar task reuse that tool instead of starting from scratch?

The loop asks Jev for a bounded next step. If no registered tool fits, a coding agent proposes and implements a narrow tool contract. CasaJev runs checks in an isolated worker, registers the versioned tool, and resumes the original task. Later tasks can select the registered tool directly. The repo includes the chat/desktop UI, task persistence, tool registry, tests, and benchmark inputs/results.

In one live example, building a new tool took 29.29 s; reusing it on a different input took 2.07 s. In a separate six-task synthetic pilot with existing tools, the Jev path had a 1.80 s median versus 13.01 s for Sol and 16.84 s for Terra through the measured integrations. These are small local measurements, not evidence that CasaJev is generally faster than GPT. Graph hints showed no measured speed benefit in that pilot.

It's Apache-2.0 and still a prototype. macOS has been exercised end to end; targeted Windows browser/desktop tests pass, but full Windows validation is pending because Docker was unavailable on the test machine. Linux end-to-end validation is also open.

Repo: https://github.com/8endit/CasaJev

I'd welcome contributors or critical feedback on independent tool verification, larger-scale tool retrieval, and a benchmark with real recurring tasks. What failure case would you test first?

— Blocboiven

How Jev fits in the loop

  1. The harness presents current state and registered tools to Jev.
  2. Jev chooses the next bounded step or indicates that a new tool is needed.
  3. A coding agent builds the narrow tool, an isolated worker checks it, and the registry stores the versioned tool.
  4. The original task resumes and later tasks can select the registered tool directly.

How to reproduce

  1. Clone https://github.com/8endit/CasaJev and follow the repository's local setup instructions.
  2. Run a task where no registered tool fits, then inspect the narrow tool contract, isolated checks, version registration and resumed task.
  3. Repeat the author's tool-build/reuse example and six-task synthetic pilot without generalizing beyond the local measurements.

Why this build matters

This is a concrete tool-learning loop: Jev decides where the agent should go next while code and sandboxed checks own implementation and verification.

Limitations

  • The post explicitly describes the project as a prototype and says the measurements are small local or synthetic tests.
  • Linux end-to-end validation remains open; full Windows validation was pending because Docker was unavailable on the test machine.
  • The author reports that graph hints showed no measured speed benefit in the six-task pilot.
  • Tool-contract or API semantic drift can still escape detection when descriptors and output schemas do not change.

Patterns