Collect action identity, outcomes, state and derived evidence locally. Raw prompts and source are not required as evidence fields.
No-progress loops
AI agents repeat work.Progress is what matters.
A practical model for detecting successful agent activity that repeats without changing observable state or producing new evidence.
- Open source
- Local first
- Shadow first
- Fails open on ambiguity
The signal
Activity is not progress.
A coding agent can execute a tool successfully and still learn nothing new. The useful question is not “did the call succeed?” but did observable state or useful evidence change?
same semantic action
+ successful outcome
+ unchanged observable workspace state
+ no new evidence
→ no-progress repetition candidateMARGINAL treats that pattern as evidence of diminishing marginal value. It does not label every retry as waste: failures, changed state, new evidence and unsupported outcomes remove or reset repetition pressure.
Why fixed token caps are not enough
A budget cap can stop a run after a threshold, but it cannot tell whether the next verification call is useful. No-progress governance is state-aware: the decision depends on what changed, what was learned and what the runtime can actually prove.
Mechanism
Observe → prove → earn.
Look for repeated successful work with unchanged state and no new evidence. Ambiguity fails open.
Only an adapter with real blocking capability and sufficient local evidence can move beyond advisory behavior.
No universal savings claim
MARGINAL measures governance overhead and workload outcomes. A historical exploratory smoke observed a token difference but did not establish causal savings; the project keeps that distinction explicit.
FAQ
Fast answers.
What is a no-progress loop in an AI agent?
Repeated work where the semantic action succeeds but observable workspace state and useful evidence do not change.
Does MARGINAL stop every repeated action?
No. Verification, failures, changed state, new evidence and ambiguous outcomes are treated conservatively. Unsupported or ambiguous cases fail open.
Is MARGINAL just a token limiter?
No. Token budgets are one cost signal; MARGINAL focuses on whether another action is expected to create enough verified value to justify its cost, latency and risk.
Try it
Observe first. Prove waste. Earn enforcement.
Install MARGINAL in Shadow Mode, inspect what your agent actually repeats, and contribute traces that make the governor harder to fool.
Deep dive: detect no-progress without reading prompts
The Research & Engineering article explains the evidence model, privacy boundary, counterexamples and why repetition alone is not enough.