Claude Code Agents Now Write Notes to Cut Context Drift
Key insights
- Anthropic's 'dreaming' feature lets Claude Code agents write structured self-notes mid-task to reduce context drift across long-horizon work.
- Nearly half of a packed London developer audience reported shipping a fully Claude-written PR in the prior week.
- A consolidation layer surfaces cross-session patterns, extending coherent agent behavior beyond single-context limits.
Why this matters
Long-horizon context drift has been the primary technical ceiling on agentic coding workflows, and a production-stage mitigation from Anthropic signals the barrier is falling faster than most engineering teams have planned for. The 'half the room raised their hands' data point is a rare, unscripted signal about actual adoption rates among self-selecting but technically serious developer audiences, not survey respondents. For technical leaders, the implication is that the internal debate about whether to allow Claude-authored PRs into production is rapidly shifting from a policy question to a workflow design question.
Summary
Anthropic's two-day Code with Claude conference in London unveiled a capability called 'dreaming': Claude Code agents write structured notes to themselves between tasks, and a consolidation layer surfaces patterns across long-horizon work to keep context coherent over time. The feature directly addresses one of the hardest unsolved problems in agentic coding — models losing track of intent and accumulating errors as task length grows.
The on-stage moment that landed hardest: an engineer asked the packed room who had shipped a PR written entirely by Claude in the past week. Nearly half the hands went up.
Essentially: (Anthropic, Claude Code) are positioning agents as autonomous contributors, not just autocomplete.
- 'Dreaming' involves mid-task self-note-writing plus a cross-session consolidation layer to reduce long-horizon drift.
- The conference framing, per MIT Technology Review, was explicitly a 'changing of the guard' in software engineering.
- The audience response signals that fully Claude-authored PRs are already a routine workflow for a significant slice of working engineers.
The shift from writing code to directing and reviewing it is no longer a prediction being made about the future; it is apparently already the present for a meaningful portion of the developer population.
Potential risks and opportunities
Risks
- Engineering orgs that accept Claude-authored PRs without updated review protocols risk shipping subtle logic errors that pass automated tests but reflect drift artifacts baked into the agent's self-notes.
- If the consolidation layer surfaces patterns across sessions by retaining note history server-side, enterprises in regulated industries (finance, healthcare) may face compliance exposure before Anthropic publishes a data-handling policy.
- Junior engineers whose workflows shift entirely to directing and reviewing Claude output before building strong independent coding foundations could create a skills-gap problem that surfaces 12-24 months from now when agentic systems require expert debugging.
Opportunities
- Code review tooling vendors (Graphite, CodeRabbit, Reviewpad) can position directly against the 'humans now review, not write' workflow Anthropic is normalizing, with agent-aware diff analysis as a differentiator.
- Enterprise AI governance platforms (Protect AI, Lakera, Apex) have a clear opening to sell session-note auditing and drift-detection layers that sit above whatever Anthropic ships natively.
- Talent platforms and bootcamps that pivot quickly to 'AI direction and review' curricula rather than traditional coding instruction can capture a cohort of developers actively looking to retrain for the workflow Anthropic described on stage.
What we don't know yet
- Whether the 'dreaming' consolidation layer operates locally on-device or routes session notes through Anthropic's infrastructure, with attendant data-privacy implications for enterprise users.
- What the failure modes of self-note consolidation look like at scale — specifically whether compounding note errors can amplify drift rather than reduce it over very long task chains.
- No public timeline was given for general availability of 'dreaming' outside the conference demo context, leaving enterprise adoption planning on hold.
Originally reported by technologyreview.com
Read the original article →Original headline: MIT Technology Review: Anthropic's Code with Claude London Conference Previews 'Dreaming' — Agents Now Write Notes to Themselves Between Tasks to Reduce Long-Horizon Drift