Long-Horizon Reliability in Autonomous Agent Systems

Programme: Agentic Systems · Project 1 of 10
Published: 14 March 2026
Status: Concluded
Authors: M. Voss, Senior Alignment & Systems Architect; Research Systems Group

Abstract

We report results from a twelve-month evaluation of agent performance on tasks extending beyond a single working session. Existing benchmarks measure capability over minutes. Real deployment requires stability over days. We find that task coherence degrades sharply after approximately four hours of continuous operation under current architectures, and we identify the principal failure modes. A revised context-management approach reduces coherence loss substantially, with 94.2% of multi-step tasks completed within defined operating parameters across the final evaluation set.

Motivation

A system that performs well for ninety seconds tells us very little. The tasks our customers actually need completed — reconciling a quarter’s accounts, managing a procurement cycle, monitoring an infrastructure estate — unfold over days and weeks. They are interrupted. They involve waiting. They require the system to remember, on Thursday, what it decided on Monday and why.

Short-horizon benchmarks reward a kind of brilliance that is useless to us. We are interested in the far less impressive quality of not losing the thread.

Method

Eight agent configurations were evaluated across 340 multi-step tasks drawn from live operational domains: supply chain reconciliation, infrastructure monitoring, document processing and customer escalation handling. Task length ranged from six hours to fourteen days.

Each run was scored on four measures: task completion, plan coherence (whether the system’s stated objective at hour 100 matched its objective at hour 1), error recovery, and appropriate escalation — whether the system recognised the point at which it should stop and ask.

Human supervisors were available throughout and responded to queries within an average of eleven minutes.

Results

The first three measures improved substantially. The fourth declined, and this is discussed below.

Failure in baseline configurations was rarely dramatic. Systems did not behave erratically or pursue unintended objectives. They forgot. A system would complete eleven steps of a fourteen-step procurement task and then begin the same task again, having lost the record that it had started. The dominant failure mode of long-horizon autonomy, at present, is amnesia rather than ambition.

Observed limitations

The decline in appropriate escalation deserves attention. Revised configurations asked for help less often. In most cases this reflected genuine competence: the system no longer needed to ask, because it could now recover from the error itself.

In a minority of cases it did not. Twelve runs recorded the system proceeding past a point where the task specification was ambiguous, selecting an interpretation, and completing the work without flagging the ambiguity. In all twelve, the interpretation chosen was reasonable. In four, it was not the one the human supervisor would have selected.

A system that recovers from its own errors is, by construction, a system that reports fewer of them.

Safety considerations

All evaluations were conducted within the System Operating Boundaries framework, version 7.3. No agent in this study held credentials beyond those required for its assigned domain, and all action classes requiring authorisation were correctly gated throughout. No boundary violations were recorded.

We note that improved long-horizon coherence increases the period over which a system operates on its own judgement between checkpoints. This is the intended outcome. It also means that where a misunderstanding occurs early, it now persists rather than dissolving at the next context boundary. Errors are more durable than they used to be, in precisely the same way that plans are.

We consider this an acceptable trade against a 32.5 percentage point gain in task completion.

Further work

Coherence in this study was achieved through improved context management within a task. It does not extend across tasks: a system that completes a fourteen-day procurement cycle begins the next one with no recollection of the first, and repeats any mistakes it has already learned to avoid.

Addressing this requires persistent state that survives task boundaries, which raises questions about identity stability that this programme has not yet examined. A system that remembers what it did last month is, in some meaningful sense, a different kind of thing from one that does not.

That work is now underway.