Executive summary
Nenya at a Glance puts the July programme on five pages. It starts from the problems every long-running agent meets: rules fall out of the context window, summaries harden tentative points into settled facts, and rules nobody wrote down get broken unnoticed. In the paper’s thirty-day model, an agent running Nenya keeps 95% of its standing commitments; unaided agents keep fewer than half. That is what lets you hand an agent a long job and trust it to finish.
Findings
- On a regulated task with four kinds of constraint, Nenya reached 1.000 fidelity. The closest alternative scored 0.886 and an unaided agent 0.429.
- An unaided model breaks an unwritten rule 55 to 61% of the time on every encounter, with no improvement. Nenya’s first encounter matches that rate; after that its violation rate stays under 1%. Over eight encounters that comes to 0.61 expected violations against 4.64.
- Adding commitments barely moves the result: fidelity is 0.995 with four running at once and 0.986 with forty-eight. An unaided agent falls from 0.429 to 0.250.
With Nenya, what you tell your agent on day one still holds on day thirty.