Performance
Judge us by how long it holds.
This page is how to judge the product of our research: one measure, tracked over time against the best alternatives, on the way to one goal. An agent that keeps its judgment, its rules and its mandate for as long as the work runs. Days now. Then months. Then for good. That is our contribution to operational alignment.
Days an agent stays reliable, with the Durable Intelligence layer
Still reliable when the 30-day test ended. Against 5.6 days for LangGraph Deep Agents with memory and checkpointing. In our evaluation, our September 2026 paper.
Days an agent stays reliable
Higher is better. Latest measurement September 2026, our September 2026 paper.
Durable Intelligence
still reliable when the 30-day test ended
LangGraph Deep Agents
memory + checkpointing
Best-practice prompting
same model, better instructions
Unaided frontier model
the model on its own
Reliable means at least 90% of the work is still correct (task fidelity), on regulated work. The count stops the first time it drops below that. In our evaluation.
Drawn to scale. Thirty is a lower bound set by the observation cap.
Scroll across to compare every milestone →
The time axis uses equal spacing between milestones; the intervals are not to scale. Each point is one product or method at the horizon it held, stamped with when we measured it. New measurements are added as they land; earlier ones keep their dates. Goals are goals, not measurements.
Work still correct on day 30
Higher is better. 6,000 runs, the same three models, blind grader.
In our evaluation. Reliable means holding at or above 90%.
| measure | unaided | best-practice prompting | LangGraph Deep Agents | Durable Intelligence |
|---|---|---|---|---|
| Work still correct ↑task fidelity, % | 41.8 | 47.6 | 63.9 | 95.2 |
| Runs that broke no hard rule ↑zero hard violations, % | 2.1 | 13.6 | 18.4 | 97.6 |
| Promises kept ↑commitment retention, % | 13.4 | 38.7 | 46.1 | 96.8 |
| Made-up outputs ↓unsupported outputs, % | 21.3 | 17.9 | 8.2 | 0.7 |
| Got stuck in a loop ↓runs entering a loop, % | 16.8 | 15.7 | 9.6 | 0.3 |
| Days reliable ↑first sustained drop below 90% | 1.8 | 2.0 | 5.6 | ≥30 |
Full scorecard, day 30. Regulated work, 6,000 runs; arrows show which way is better. Models: GPT-5.5, Opus 4.6, Gemini 3 Pro, the same three in every column. Per model with the layer: 95.6, 96.1, 93.9. Per workload spread 1.1 points. In our evaluation.
How much a person has to babysit it
42-day programs, 384 of them. Durable Intelligence versus a frontier agent.
Runs on its own for, before a person has to step in
Longer is better. 14.8× longer between interventions. About 4.6 days. In our evaluation.
Human effort to work with the agent, per day
Less is better. About a third of the human time. In our evaluation.
Accepted work packages
Most work packages still fail. In our evaluation.
Medians over the same 384 programs: 109.6 vs 7.4 hours between interventions; 13.7 vs 41.6 hours of human time per 42-day program, shown here divided by 42. Accepted work rose to 24.3%; most work packages still fail. In our evaluation.
What this means
Task fidelity by day, in our evaluation. Hollow marks: the reliable horizon, 1.8 and 5.6 days.
The gap is not fixed. It widens with the horizon and with how strictly the work is judged. At a day or two LangGraph Deep Agents and best-practice prompting look close; by day 30, and on the measures that count a run as failed if a single rule is broken, the distance is the whole point.
The base models did not change. Across every row it is the same three frontier models doing the work. Only the system around them changed, which is why the comparison is fair and why the difference is attributable to the layer, not to a stronger model.
In plain terms, holding buys a team three things: fewer rebuilds when an agent quietly loses the plot, less time pulled in to supervise it, and more of the finished work accepted instead of thrown out. That is the difference between a demo and a worker.
Not solved yet
Read this before you believe the rest.
Work packages accepted as done 24.3%, up from 16.2%
Each mark is one tenth of a percentage point. The unfilled marks are the work still failing.
- Accepted work 24.3%
- up from 16.2%; most work packages still fail
- 109.6 hours before a manual rebuild
- about four and a half days, not 42
- Day-42 survival 71.4%
- up from 31.8%
- Horizons past 30 days
- unmeasured; the 30-day figure is an observation cap
- Our grading
- blind grader, fixed rubric, but ours; independent replication is the next step
- Benchmark-adjacent comparisons
- the Remote Labor Index figure uses a dated reference score, not the official leaderboard
What it is
The Durable Intelligence layer, the product of the research.
- Predictive-state control
- estimates the agent's operative state and applies targeted state changes, so a rule set on day one still governs decision three hundred
- Drift correction and verified compaction
- compaction keeps what governs; corrections propagate to the decisions that depend on them (causal repair)
- Execution renewal
- recovery through restarts, tool failures and model succession without losing the mandate
- Effect authorization
- a separately governed, non-bypassable boundary for anything that touches the outside world
- Model-agnostic by construction
- the current hosted-model interface uses no base-model activations, gradients, or fine-tuning; it acts through inference inputs and exposed message and tool boundaries
- Deployment
- one workflow first, a two-week look against your own baseline, then wider; nothing retrained, nothing migrated
- Measured on
- legal review, M&A diligence, Medicare coverage, grant preparation, government compliance
How we measured
Thirty days, one hundred runs per cell, blind grader, rubric fixed in advance.
Six thousand runs: four configurations, three frontier models, five regulated workloads, one hundred runs each, thirty days, a blind grader and a rubric fixed in advance. A separate 384-program, 42-day lifecycle evaluation with recurring maintenance, delayed obligations and changing evidence. The models were GPT-5.5, Opus 4.6 and Gemini 3 Pro, held fixed across configurations. The grading is ours for now: a blind grader against a rubric fixed in advance; independent replication is the next step, and the protocol is in our September 2026 paper, Dawn of Durable Intelligence, so anyone can run it. This is a product page; the paper is the record.