Agent Decision Layer
yankis@yankis0x
9:29This project implements a decision layer that sits before expensive agent actions, determining whether to act, escalate, or involve a human based on confidence scores.
0.67 confidence. Your agent acted on it anyway.
99% of agent setups never set a bar. The 1% know 0.67 is the do-not-act band.
A decision layer sits in front of the expensive steps - browser, deep research, retries, subagents - and decides whether they happen at all.
setup is a clone and a config :
state -> one Jev call -> policy -> act, surface, or escalate
step 1 -> export your TypeSafe key in your shell profile, not in the session
step 2 -> clone muse-jev-playbook, venv, requirements, copy config.example.yaml
step 3 -> build a compact state before every fork: goal, kind, cache hints, error history, constraints
step 4 -> ask all three primitives in ONE parallel call: choice, score, noul, each with a confidence 0-1
step 5 -> branch on policy, never on vibes: 0.80 and up act, 0.50 to 0.79 go to a human, below 0.50 escalate
step 6 -> run it in shadow, read the log, flip to active only when the log agrees with you
step 7 -> use the recipes instead of writing prompts: triage, rank options, act-or-wait, retry gate, research cap, approval gate
the result, from the playbook's own flight case:
- 92 days of fares collected, range $698 to $895, median $851
- narrowed to 8 October days before any deep dive
- $153 under the median, 18%
the part the setup posts leave out:
that same case asked for the single best day and got 0.67. surface band, not act band. on the round trip, book-now came back 0.79, one hundredth under the bar.
so the layer hands you the month at 1.0 and gives the day back to a human. that is the policy working, not failing.
a decision layer does not make an agent smarter. it makes it STOP in the right places. on a real bill, that is worth more.
Related builds
