IndyDevDan’s 35-minute walkthrough builds from simple decisions to Jev tools inside a Pi coding agent. I read the complete transcript and inspected the companion lab at revision 777adaf47d37ae0553220d35b2f15b3a3a063305. 201 offline tests passed; 10 live tests were skipped. I did not reproduce the hosted decisions, Pi sessions, or speed/cost comparisons; this is not a security audit.

01:03–14:30 — Define the question; keep the policy in code

  • Start with yes/no decisions, then declared-choice ticket triage and multiple questions about one state.
  • Score separate factors—severity, frustration, review risk—and combine them with explicit application weights.
  • Use uncertain decisions to request review rather than treating every typed answer as permission.
  • Route to the right model, agent harness, or workflow; a classification does not establish that the chosen agent can finish.

14:35–19:53 — Hooks have boundaries

  • Pre-tool hooks demonstrate blocking destructive commands and writes; turn-end hooks judge when work has changed enough to compact.
  • The author’s demo uses 6k/10k/14k-token compaction thresholds to make transitions visible, not to establish production defaults.
  • The repository documents a write-gate bypass through bash; telling the agent not to work around a block is an instruction, not isolation.
  • The inspected guard catches and logs scoring errors without returning a block: a failed check can leave execution allowed.

20:06–28:42 — Scout files outside the coding agent’s context

  • A tool reads files in ordinary code, sends their content to Jev, and returns only answers to the main agent.
  • Extend that to paths, globs, or recursive scouting; then read relevant files normally before editing or quoting them.
  • The inspected batch implementation bounds concurrency and reports skipped files and failed calls; an incomplete result is not an all-clear.
  • “The agent did not read the file” does not mean the file stayed local: hosted Jev receives it.

29:34–32:28 — Let the agent ask, not certify itself

  • The final tool lets the coding agent write Jev questions over state, files, and command output.
  • The demo classifies a failing rounding test, fixes code, and asks for another judgment after the tests run.
  • That tool gates its own command path; the repository explicitly says the agent’s separate bash tool is not gated at this level.
  • Preserve exit codes, actual assertions, permission controls, and human review; a model’s “tests passed” judgment is not the test result.

What this establishes

The presenter shows concrete compositions, not a controlled benchmark. His subsecond responses and large price multipliers are presenter claims, not independently reproduced results; price-table extrapolations omit important end-to-end workload differences and caching.

He repeatedly calls scores “confidence intervals.” They are model-returned probabilities or confidence values—not statistical intervals, and not evidence of calibration on these tasks. Typed output is not a correct decision.

“I am using Jev as an extension of my agent. It’s not a replacement.” — IndyDevDan, approximately 21:48