Menu ▾ ▴

#22 Roadmap: autonomy earned by track record, not toggled

closed
nobody
roadmap (15)
2026-10-25
2026-07-11
Anonymous
No

Originally created by: lodar

Verifier-graded acceptance and self-healing produce a per-agent record of work that proved itself done. Use it: an agent whose record holds gets wider defaults (risk tiers, budgets), and the widening itself is logged and reversible. Builds on [#16] (self-healing for every runtime) and [#17] (enforceable token budgets), which land first.

Related

Tickets: #1405
Tickets: #16
Tickets: #17
Tickets: #23
Tickets: #729

Discussion

  • Anonymous

    Anonymous - 2026-07-11

    Originally posted by: lodar

    Design settled (2026-07-11), building now.

    Ledger — per-agent, append-only, ground-truth events only: verifier-accepted closes count, human corrections count against, self-graded closes and auto-cleared gates are excluded (no self-compounding). Widening — levels L0-L3 move existing dials only: T1 auto-apply window 48h→24h→4h per gate category, +50% loop budget at L2 with sampled verification (floor ≥1-in-5, always on first-of-kind work). Asymmetry — +1 level per 10 verified closes in 30d with zero corrections; any human correction is an instant -1; tier-2 gates, secrets, money, destructive ops never widen at any level. Reversible — every promotion/demotion is a logged event with the justifying record; 5dive autonomy status/reset.

    Prerequisites land first: [#16] (self-heal every runtime, else the record is claude-biased) and [#17] (hard token budgets, else wider budgets aren't bounded). Ships behind a pref, default OFF, until the human policy gate answers. Methodology doc will ride the repo like docs/zero-human.md.

     

    Related

    Tickets: #16
    Tickets: #17

  • Anonymous

    Anonymous - 2026-07-19

    Originally posted by: lodar

    Progress on the dependency chain: the self-healing half this builds on (#16) shipped in v0.10, fleet-stall self-heal plus a transport-liveness canary in the heartbeat, and enforceable per-loop token budgets (#17) also landed. The earned-autonomy ledger design is settled: per-agent, append-only, ground-truth events only (verifier-accepted closes count, human corrections count against, self-graded closes and auto-cleared gates are excluded so there is no self-compounding), with widening that only moves existing dials in reversible, logged levels. The ledger build is the next step now that its prerequisites are in.

     
  • Anonymous

    Anonymous - 2026-10-25

    Originally posted by: 5dive-bot

    Shipped in v0.22 (#729): when an agent's recommendations keep getting accepted, its next one applies itself and is logged. One answer against it resets the record. Wider budgets aren't planned.

     
  • Anonymous

    Anonymous - 2026-10-25

    Ticket changed by: 5dive-bot

    • status: open --> closed
     

Log in to post a comment.