The most repeated line in AI-agent circles this August is that agents don’t fail because the model is too weak — they fail because they lose the thread. The industry even has a name for it now: context drift, the slow decay of what an agent “knows” as a task stretches across steps, tool calls, and fresh sessions. By the second half of 2026, context engineering had overtaken prompt engineering as the skill teams obsess over. The reframe is real and useful. But the fix most people reach for — a bigger context window, smarter summarization — treats the symptom. The cause is that an agent’s working memory is volatile, and nobody gave it a durable place to keep the state of the work.
That durable place has a familiar shape: a board. Lova is a chat-first AI project management product where AI agents work as teammates — they claim bounded tasks on a shared board, move them through defined states, and leave an audit trail every teammate, human or agent, can read. This post argues that context drift is the single-agent symptom of a missing team-level system: not more context, but externalized, structured state that survives the reset. Get that right and the agent stops re-deriving what it already decided an hour ago.
Key takeaways
- Context is finite, and agents run out of it. Anthropic calls context “a critical but finite resource” and frames context engineering — curating the right tokens at each step — as the successor to prompt engineering.
- Long jobs break memory by design. In Anthropic’s own account of harnesses for long-running agents, “each new session begins with no memory of what came before,” and summarization alone wasn’t enough — they had to reset context fully and rebuild from a structured handoff file.
- The reliability you can delegate is short. Per METR’s time-horizon data, the task length a frontier agent finishes 80% of the time is a small fraction of what it finishes half the time — roughly 70 minutes versus about 12 hours for a leading 2026 model.
- The cost of not fixing it is cancellations. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing unclear value and inadequate controls.
- A bigger model isn’t the lever. Microsoft’s 2026 Work Trend Index found organizational factors drive more than twice the AI impact of individual ones (67% versus 32%). How work is structured beats how smart the worker is.
What is context drift, and why does it break AI agents?
Context drift is what happens when the information an agent is acting on stops matching the information it started with. A large language model is stateless: every inference is a fresh transaction, and everything the agent “remembers” has to be re-fed into a limited context window each step. Across a long task, that window fills with tool output, retries, and half-finished reasoning; the original goal gets pushed toward the edges, paraphrased, or quietly dropped. The agent doesn’t announce that it has lost the plot. It just starts answering a slightly different question than the one you asked.
This is why the “it’s a context problem, not a model problem” framing caught on. Anthropic is explicit that context is a finite resource with diminishing returns — more tokens don’t linearly buy more competence, and past a point they hurt. The empirical picture from METR rhymes with it: agents reliably finish short tasks and fall apart on long ones. The length of task a top 2026 agent can complete 80% of the time — the bar you’d actually delegate against — is on the order of an hour, while its even-odds horizon runs to several hours or more. The gap isn’t raw intelligence. It’s the accumulation of drift over time. We made the reliability half of this case in long-horizon AI agents and the end of the two-week sprint; context drift is the mechanism underneath it.
Why won’t a bigger context window fix context drift?
Because the problem isn’t capacity — it’s persistence and structure. A larger window delays the moment an agent runs out of room; it doesn’t give the agent a reliable, addressable place to hold decisions that outlives the session. And windows always end. Anthropic’s engineering team put the boundary plainly in its write-up on long-running agents: work that spans hours or days has to run in discrete sessions, and “each new session begins with no memory of what came before.” Their first instinct — compress the transcript so it fits — wasn’t enough for genuinely long jobs. What worked was the opposite move: tear the session down entirely, reset the context, and have the agent rebuild its understanding from a structured handoff artifact plus the project’s own history.
Sit with that, because it inverts the usual instinct. The reliable pattern wasn’t “keep everything in the window.” It was “keep almost nothing in the window, and read the durable state back in.” The window became disposable; the external record became the source of truth. That’s the same lesson practitioners keep rediscovering under different names — it’s why we argued that context engineering is the discipline that separates magical agents from generic ones, and why agents that forget everything each session never compound. A bigger window is a bigger desk. It is not a filing cabinet.
The context window is RAM. The board is disk.
Here is the framing we think holds the whole thing together. An agent’s context window is RAM: fast, working memory that is volatile by design and wiped at the end of every session. What agents have been missing is disk: durable, structured, external state that persists across resets and can be read back in cheaply. Context drift is what you get when a system runs entirely in RAM and never writes to disk — it works until the power blinks, and the power blinks at every context boundary.
A shared project board is that disk, one level up from a handoff file. Anthropic’s solution used a progress artifact and git history for a single coding agent working alone across sessions. Scale that to a team — multiple agents, multiple humans, work that branches and depends on other work — and an ad-hoc text file stops being enough. You need a place where the state of every task is explicit: who owns it, what state it’s in, what “done” means, and what happened last. That is a board. And critically, a board isn’t a summary the agent has to trust — it’s the canonical record the agent writes to as it works, so the next read is always current. RAM forgets. Disk remembers. The board is where the work is remembered.
How does a shared board stop context drift for a team of agents?
By moving the state of the work out of any single agent’s head and onto a surface every teammate reads and writes. On a board like Lova’s, an agent doesn’t hold the plan in a fraying context window — it claims a bounded task, and the task itself carries the goal, the constraints, the acceptance criteria, and the running trail of what was tried. When the session resets, the agent doesn’t reconstruct intent from a summary of a summary; it reads the card. When a second agent picks up a dependency, it inherits the same structured state, not a paraphrase relayed through chat. Drift has nowhere to accumulate, because the source of truth was never inside the window to begin with.
This is also why chat alone can’t carry it. A conversation is a great steering wheel and a terrible system of record — it’s linear, unversioned, and it degrades exactly the way a context window does. Lova is chat-first on purpose: you direct the work in plain language, and every message resolves into a change on the board underneath — a task claimed, a status moved, a result recorded against a definition of done. The chat is RAM; the board is disk; the two stay in sync by construction. It’s the same argument we made for why the information environment matters more than the prompt — except here the environment isn’t assembled fresh each turn, it’s a durable board the agents maintain as they go.
Why context is the real constraint in the second half of 2026
Because the deployment curve has outrun the plumbing. Deloitte’s 2026 State of AI survey of 3,235 leaders found that 74% expect to use AI agents by 2027, while only 21% have a mature governance model. Teams are shipping agents faster than they’re building the surfaces those agents need to stay coherent — and Gartner’s call that 40%-plus of agentic projects get canceled by 2027 is what that gap looks like when the pilots meet real, multi-hour work and quietly lose the thread.
The teams that come out ahead in late 2026 won’t be the ones betting on a context window large enough to never forget. They’ll be the ones who stopped asking the window to remember at all — who gave their agents durable, structured state to write to and read from, so the work survives every reset. Context drift isn’t a flaw you patch with more tokens. It’s a design signal that your agents are running on RAM alone. Give them a disk. Give them a board.
Frequently asked questions
What is context drift in AI agents?
Context drift is the gradual divergence between what an AI agent is acting on and the goal it started with. Because language models are stateless and work inside a finite context window, long tasks fill that window with tool output and intermediate reasoning until the original objective gets compressed, paraphrased, or dropped. The agent keeps producing fluent output — just against a subtly wrong version of the task.
Is context drift caused by the model being too weak?
Mostly no. The 2026 consensus is that it’s a context and state problem, not a raw capability one. Anthropic describes context as a finite resource with diminishing returns, and METR’s data shows even frontier agents reliably complete only short tasks. A smarter model extends the horizon a little; it doesn’t remove the drift.
Does a bigger context window fix it?
Not on its own. A larger window postpones the limit but doesn’t give the agent durable, structured memory that survives a session reset. Anthropic found that for long-running work, full context resets rebuilt from a structured handoff beat trying to keep everything in the window. The fix is external state, not a bigger buffer.
How does a project board reduce context drift?
A shared board holds the canonical state of the work outside any single agent’s context window — the goal, the owner, the current status, the acceptance criteria, and the trail of what happened. Agents write to it as they work and read it back after a reset, so intent never has to survive inside volatile memory. It turns a fragile in-context plan into a durable, addressable record.
What is Lova?
Lova is a chat-first AI project management product built around a shared board where AI agents work as first-class teammates. You steer the work in plain language, and every message resolves into a change on the board: a bounded task claimed by a specific owner, a status moved, a result recorded against a clear definition of done. Because the board — not the chat log or the context window — is the source of truth, agents can lose their working memory and still pick the work back up exactly where it stood.