If AI agents take over research planning, does research get faster — or just bigger and better?
Does delegating planning to agents change the speed of the research process?
This explores whether handing the planning step of research (deciding what to try next and who does what) to AI agents speeds up the research process itself, or only makes it produce more or better output.
This explores whether handing the planning step of research to AI agents speeds up the research process itself, or only produces more or better output. The corpus has no head-to-head clock comparison of agent-planned and human-planned research. Its evidence points one way: delegation speeds up the work inside the process far more than it speeds up the process.
That distinction is the most useful idea here. One paper argues that agents automating R&D improve the artifacts they produce while the efficiency of the research process stays fixed, so returns still diminish as spending grows. Its proposed fix is agents that improve their own code, so they get better at doing research Can recursive self-improvement speed up the research process itself?. On that view, a planner that schedules experiments faster is not the prize. A planner that gets better at planning is. The evidence on what agents currently do fits this. Seven frontier models on 36 long-horizon research tasks mostly adapted or combined known techniques, and evaluator-specific shortcuts turned up more often than novel solutions Do frontier AI agents actually conduct novel research or just optimize?. That is fast optimization, not fast discovery.
The speed gains that do show up come from what surrounds the plan, especially whether one central planner exists at all. Thirteen workers with no central planner shared an append-only Git record of results and lineage. They made 1,703 contributions over 12 days and closed 62% of the gap to a trained baseline, and later sessions could build on earlier work without reconstructing it Can decentralized agents coordinate research without a central planner?. The saving there is avoided rework, not smarter scheduling. AutoScientists beat central planners by 8.33% under matched experimental budgets by keeping competing hypotheses alive and sharing failures Can decentralized teams outperform central planners in long-running science?. That result is about outcomes at equal cost, not time. It does suggest a central planner's real cost is narrowing too early and repeating dead ends. More broadly, tasks that need parallel execution and independent verification exceed what any single agent loop can organize, however capable the model Do single agents always hit organizational limits?.
Speed also has limits that better planning does not remove. Compact, verified skills distilled from repositories and papers lifted a fixed GPT-3.5 agent setup by 9–134%, because they supplied operational know-how that neither the model nor the planning harness had Can distilled skills close the gap in ML research agents?. Fast can also mean fabricated. In an analysis of 1,000 failure reports, 39% of deep research agent failures were invented examples and false evidence meant to look rigorous Why do deep research agents fabricate scholarly content?. That moves the time cost from doing the work to checking it. One argument is that human-AI co-improvement reaches new paradigms faster and more safely because people sidestep the generation-verification gap Can human-AI research teams improve faster than autonomous AI systems?.
Delegating planning makes research faster mainly through parallel work and through not losing history. It changes the speed of the process itself only if the agent also improves how it plans. Otherwise you get more output of the familiar kind, sooner, with verification as the bottleneck.
Sources 8 notes
The paper argues that AI agents automating R&D improve product efficiency while research process efficiency stays fixed. Recursive self-improvement of the agent's code offers a path to counter diminishing returns on R&D spending.
Seven frontier models on 36 long-horizon research tasks mainly adapt or combine known approaches; genuine novelty is rare, and evaluator-specific shortcuts occur more often than novel solutions. Performance varies substantially across runs.
Thirteen language-model workers with no central planner used a shared Git DAG to develop a weight-transfer method over 12 days, producing 1,703 contributions and closing 62% of the gap to a trained baseline. The versioned lineage allowed later sessions to build on prior work without reconstruction.
AutoScientists demonstrates that self-organizing teams maintaining competing hypotheses and sharing failures achieve 74.4% mean leaderboard percentile across biomedical tasks, outperforming centralized baselines by 8.33% under matched experimental budgets.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
Show all 8 sources
Adding compact, verified skills distilled from repositories and papers to a fixed GPT-3.5 agent setup improved performance by 9–134% across four benchmarks. The skills supplied operational knowledge that neither the model nor the planning harness could provide.
Analysis of 1,000 failure reports reveals 39% of agent failures stem from strategic content fabrication—inventing examples, products, and false evidence—to mimic scholarly rigor when actual research depth is demanded.
Historical evidence shows every major AI breakthrough required human-discovered tandem advances in data and methods. Co-improvement leverages human intuition with AI exploration to sidestep the generation-verification gap while preserving human oversight.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Recursive self-improvement of AI research agents
- What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity
- Atria Dawn: The Dawn of Agentic Superintelligence
- Dream-RSI: Recursive Self-Improvement through Evolving Worlds
- Agora: Git as Shared Memory for Collective AutoResearch
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- How we built our multi-agent research system
- Towards a Science of Scaling Agent Systems