Context drift
Arthaud
·
Mar 2026
Since adopting agentic engineering, we 4x'd our output, but the cost of coordination exploded.
We call it context drift: the growing gap between what different people on a team believe is true.
Like chinese whispers, drift has always been there. Between humans, across functions, through every handoff.
Rituals were built to limit it: roadmaps, standups, sprints. Agents don't create drift. They just make it compound faster. A 1% misalignment on Monday is a shipped feature nobody needed by Wednesday.
Code has versioning, limiting drift
In three years, code went from autocomplete to synchronous agents to async agents. But code had 50 years of version control infrastructure to absorb the shock. Context lives in one place: the repo. There's branches and merges for divergence. A compiler that rejects incoherence. And even with all that, teams still accumulate tech debt, still ship bugs, still drift from intent. Code barely holds it together.
Knowledge work doesn't have the same guardrails
Knowledge work is at the start of that same automation arc, and it has nothing. Context is scattered across Slack, Notion, Linear, CRM, email. Owned by different people, updated at different times. No merge protocol. No canonical "main." And the most critical context, intent, priorities, trade-offs, never gets written down at all. It lives in people's heads, in hallway decisions, in assumptions nobody thought to document. When two people operate on different assumptions, nothing flags the conflict. It just ships.
Before agents, teams stayed aligned by accident. Things moved slowly enough that email, meetings and informal chats kept everyone synced. Remove the friction, you remove the synchronization. Add agents, and models write faster than humans can read. Decisions multiply. Understanding doesn't. That's worse than being slow everywhere.
Drift isn't a transition problem. It's the new default.
Alignment used to be a byproduct of a slower pace of work. Now it actively costs energy.
Without a source of truth, agents don't just automate the busy work. They create more of it downstream. Once you realise the context they used for a decision was stale, you cannot rewind a business decision like you rebase code.
Speed was never the bottleneck. Shared understanding was and it still is.
