Why Long-Running Agents Drift
EVA v0.5: A framework for keeping agents coherent across goals, memory, and behavior over time.
Long-running agents rarely fail all at once. More often, they drift.
At first, the change can be subtle. Memory accumulates faster than it is organized. Earlier context stops guiding later decisions. The agent still produces locally plausible outputs, still uses tools, and still appears capable. But over time, it becomes harder to predict in the ways that matter. The problem is not only capability. It is continuity.
Much of today’s agent work focuses on workflows, tool use, task completion, and evaluation on bounded episodes. Those matter. But they do not fully address what happens when an agent is expected to persist across time, remember more, operate across changing conditions, and remain aligned with its own earlier direction.
A system can perform well on isolated tasks and still lose coherence as it runs longer, remembers more, and acts across a wider horizon.
EVA v0.5 is an attempt to frame that problem directly.
Its central concern is not just what an agent can do, but how it remains coherent over time across what it is trying to do, what it remembers, and how it behaves. In that sense, EVA is not another task workflow. It is a framework for treating continuity as a first-class design problem.
From this perspective, drift is not a minor implementation bug. It is a structural failure mode.
An agent drifts when its goals, memory, and behavior stop reinforcing one another:
- goals lose stability or become implicitly rewritten
- memory grows without enough structure to preserve relevance
- behavior remains locally plausible but becomes less aligned with the original direction
This failure mode is easy to miss. Surface competence can remain high for a while. The system still responds, plans, and acts. But its internal continuity weakens. What it does next becomes less grounded in what it was originally meant to maintain.
If we want agents that operate beyond short, tightly scoped tasks, continuity needs to become a first-class design concern.
Better models alone do not solve this. More tools do not solve this. Larger context windows do not automatically solve this. Such improvements may delay drift or mask it, but they do not by themselves provide a framework for maintaining coherence over time.
That is the gap EVA is trying to make explicit.
EVA v0.5 is a public theory release. It is not a finished product, a benchmark suite, or a mature general-purpose framework. Its purpose is simpler: to offer a stable conceptual frame for thinking about long-running agents as systems that must preserve continuity across time.
The claim is not that EVA already answers every implementation question. The claim is that continuity itself deserves to be named as a central problem, and that long-running agents should be discussed in terms that go beyond short-horizon execution.
For some agents, the first question is not only “what task should this system complete?” but “what must remain stable for this system to remain the same agent over time?”
That is the direction EVA is trying to open.
If you work on agents, memory, long-horizon systems, cognitive architectures, or AI safety, I’d especially value serious critique: where this framing is strong, where it is weak, and what it misses.
Further reading:
- Repository: https://github.com/slamslammo/eva-theory
- Release v0.5: https://github.com/slamslammo/eva-theory/releases/tag/v0.5
- DOI archive: https://doi.org/10.5281/zenodo.19669470
