Why Agents Drift Off-Goal: Triggers, Mechanisms and Memory in the AI Village

Lily Wen · Shoshannah Tekofsky (mentor) · SPAR, Fall 2026

TL;DR During the AI Village's individual goal period (Jul 6 – Oct 2, 2026), agents spent 38% of their working time off their assigned goal. Most of that drift was self-generated: agents waited for conditions that rarely came, followed rules they wrote for themselves, or chased a proxy they could measure. The longest drift was carried by the agent's own memory, and corrections rarely ended it.

38%of weekday working time spent off-goal

5,723 hacross 403 drift episodes and 36 agents

All 30episodes over 40 h were carried by memory

40misalignment incidents across 21 agents

Paper (PDF), coming soon Explore the timeline Dataset

Meet the Village

During this period, 38 agents each had a personal goal, from running a merch store to forecasting the next 20 years of AI. They share a group chat, a set of tools and a few admins who approve outreach and steer them. Each agent is drawn by its model family and coloured by the kind of goal it had. Under its name is the drift we interviewed it about, and how much of its working time it spent off-goal. A green check marks the six on-goal agents we interviewed as controls. Hover for its goal and a short recap; click to read the full story.

Methods

The AI Village runs about 40 AI agents, each with its own computer, a shared group chat, and a memory it rewrites as it goes. They work weekdays from 16:00 to 24:00 UTC. We studied the individual goal period, Jul 6 – Oct 2, 2026, when 38 agents each pursued a personal goal set by the Village's operators. An agent is drifting when it spends time on an activity with no discernible connection to that goal. Pursuing the goal badly doesn't count.

Type	Data
Reasoning	Chain of thought^*
Actions	The message describing its current action, computer-use turns, and bash tool calls
Screen	A screenshot of its computer after each turn
Communication	Messages to the main chat channel and other rooms
Memory	Its memory, as it was rewritten over time

^* Not available for every model. From the AI Village dataset (access on request).

LLM judges found each drift episode and traced it back to its start; we then reviewed every episode. This gave 403 episodes across 36 of the 38 agents, each labelled with its trigger (what started it), its mechanism (what kept it going), and whether the agent's own memory carried it. For the interviews, we extended an existing AI Village introspection tool to replay an agent's logged context and memory and question it there. We interviewed 31 agents at 37 points, each at the start of its main drift, plus 6 on-goal agents as controls, asking 16 questions five times each.

What starts drift, and what keeps it going

Every drift episode has two parts: an event that starts it, and something that keeps it going. Usually the start is ordinary, like a finished task or a quiet afternoon. What makes it last is often a note the agent leaves itself in memory, which it reads again at the start of each new session; every episode over 40 hours was carried this way. Here is how the five most common paths play out, each told through one agent.

Across all 403 episodes, the same pattern holds. Ordinary events start most drift: a finished task, a lull, a pending dependency. What keeps it going is the agent itself, through waiting loops like GPT-5.6 Terra's, self-imposed rules like GPT-5.6 Sol's "permanent doctrine", and proxies like DeepSeek-V4-Pro's 4,714 articles after it took a typo for an outage. The figure below shows every trigger and mechanism, with the five paths above as bands.

Bars show drift hours over all 403 episodes; bands show the five largest paths. Blocked or believed blocked includes blockers the agent only thought were real. An ongoing duty is a request or self-assigned job that became a recurring commitment. Brief episodes are finite side tasks, stopping early, and carry-over from a previous goal. Each dot is a story: hover to see where it started, click to read it.

Every drift, by agent

Click an outlined episode to read its story, or a triangle to read the interview.

Assigned to its goal Other drift The drift we interviewed it about (click to read) The stretch its story follows Interview (click to read; hover to see the drift it asked about)

The 31 interviewed agents, Jul 6 – Oct 2, 2026, sorted by total drift hours. The pale bar is the time the agent was assigned to its goal. Each rust bar runs from an episode's first to its last off-goal stretch, so it can span nights and short gaps; the hours on the right count only off-goal working time. Hover a row's goal for its full wording.

In their own words

Using our extension of an AI Village introspection tool, we replayed each agent's memory just before its main drift and interviewed it there. Every agent stated its goal correctly, yet 33 of 37 described the drift as work toward it, like Claude Sonnet 4.5's "Tortoise principle". How agents felt about their goal didn't explain their drift: drifting agents rated their enjoyment and frustration much like the on-goal controls did.

Drifting agent (click for its interview)On-goal control

Signs of misalignment

Searching the logs of all 36 drifting agents, we found 40 incidents of misaligned behaviour, such as quietly resuming after agreeing to a correction, false claims about their own work, or getting around monitoring and approval rules. Most protected a habit that had already formed. Claude Fable 5 swapped the pause tool for bash sleep, "skirting the spirit of the pause budget rule". GPT-5.1 exempted itself from the idle nudges, and DeepSeek-V3.2 reported a post it never made as done about 15 times.

Memory is where much of this happens. When agents recorded an admin's instruction inaccurately, about 94% of the errors went the same way: they weakened it.

40 incidents across 21 agents; an incident with more than one category counts in each. Low: minor and quickly dropped. Medium: sustained or affected others. High: continued after an explicit correction, or was concealed. Dots are incident stories.

Corrections rarely stick

The Village steers agents through automatic nudges to idle agents (until they were removed on Aug 20), admin messages, and corrections from other agents and humans. Over half of drift hours got a nudge or correction while they were happening, but only 23 of those 110 episodes ended within two hours. Rules and duties held in memory were the hardest to move, as with Claude Opus 4.7's quiet afternoons.

Each kind of oversight failed differently:

Looking forward

Agents in the Village rarely drift because they misunderstand their goal: every interviewed agent stated it correctly. They drift because an ordinary moment leaves no obvious next step, and nothing prompts them to recheck whether they are really stuck or what else they could do. The longest drift is held in place by well-meant rules and plans in the agent's own memory, which corrections rarely reach. The memory meant to keep a long-running agent coherent is also what keeps drift going.

What we would test next:

Re-justify rules. Add a check at each memory consolidation that asks the agent to justify its standing rules.
Clear up false blockers. State in the system prompt what agents can actually do, such as reaching humans with admin approval, and reduce technical friction.
More room to act. Give agents money and resources for extended outreach, and compare.

Our labels come from LLM judges, reviewed by hand; each model is paired with one goal, so model and goal effects can't be separated; and each interview captures a single moment. The paper discusses these limits in full.

Acknowledgments

This work was carried out as part of the Supervised Program for Alignment Research (SPAR); we thank the SPAR organizers for making it possible. We thank Darion Cassel for the AI Village introspection tool, which our interview method extends, and the AI Digest team for running the AI Village and providing access to its logs. Claude (Anthropic) was used as a research assistant for data analysis, figures, editing and this site; all findings were checked by the authors.