A long-running agent takes on obligations faster than any fixed prompt can hold them, and keeping old work valid draws on the same budget as doing new work. We asked whether an agent can carry that growing load for a month without losing coherence.
We built a thirty-day model of five professional fields: legal, M&A, Medicare, grants and government. Because staying coherent costs compute, we measure the productive work an agent sustains once that cost is paid.
What we found
- The system kept 95% task fidelity at thirty days, against 42% for unaided agents. Fidelity barely moved across the month, from 98% at five days, and more compute raised it towards 99% and 97%. Unaided agents managed 53% at five days.
- Restating every rule at every step, the strongest conventional approach, reached 77% at five days and 48% at thirty. The system led it by more than fifteen points at both.
- On a live model, an instruction at risk of being lost held every time with the system in place, against one time in four without it.
- A model’s own report of its compliance proved unreliable: decoy rules fooled its account of which rules still held.
- Facts with a clear label stay intact deep into a long context. Running totals drift steadily as the work goes on.