Plan 0020
Agents should feel the cost of being stuck
Fable 5 sessions sometimes grind a task for two hours that should take twenty minutes, repeating a failing approach with small variations. Hypothesis: actions are painless. Make cost and stall legible in-session, prescribe an escalation ladder, and gate the method on an A/B eval against SOTA models.
- Status
- Input wanted RFC
- Ambition
- 6 changes how the harness shapes agent behavior, not just what tools exist
- Impact
- 8 two-hour grinds that should be twenty minutes burn the scarcest resource
- Effort
- 5 stall detection, budget signals, and an escalation ladder are small mechanisms
- Risk
- 4 a bad signal nags; the eval gate keeps a net-negative method from shipping
- Maturity
- 1 hypothesis stage, drafted from observed sessions
- Leverage
- 8 every agent session on every task inherits better stuck-behavior
- Taste
- 8 what pain should feel like to a model is exactly the judgment part
- Authors
andrewgazelka- Created
- Updated
- Tracking issue
- -
- Supersedes
- -
- Superseded by
- -
Summary
Watch enough sessions and the failure mode is unmistakable: the agent is not blocked, it is stuck, re-running a failing approach with small variations for two hours on a task that should take twenty minutes. A human in that loop feels pain: elapsed time, mounting cost, the embarrassment of the fifth identical error. The agent feels none of it. Every action costs the same nothing, so nothing ever forces the question “is this approach working?”
Proposal
Make the cost legible, then prescribe what to do when it registers:
- Stall detection: N attempts with no state change injects an explicit notice into context: what was tried, how long, what has not changed.
- Budget signals: elapsed wall-clock and spend, visible in-session, against a stated estimate for the task.
- Escalation ladder: on a stall the agent must pick a rung, not retry: restate the goal, change strategy entirely, widen context, delegate a subproblem, or surface the blocker with evidence.
Eval gate
The method ships only if it wins: an A/B suite of stuck-prone tasks run against SOTA models (Fable 5 and peers), measuring time-to-completion, success rate, and quality of giving up (does the blocked report name the real blocker). If the signals do not move those numbers, they come out.
Drafted by Claude Code (Fable) from operator observations; comments wanted, hence the RFC flag.