Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note4 min read

Project Fetch: Phase Two

Source

Anthropic Research

anthropic.com/research/project-fetch-phase-two

Read the source

Source

  • Source name: Anthropic Research
  • Source kind: blog (industry — first-party vendor research)
  • Authors / organization: Michael Ilie, C. Daniel Freeman, Kevin K. Troy / Anthropic Frontier Red Team
  • Venue: Anthropic Research Blog
  • Published: 2026-06-18
  • Collected: 2026-06-22
  • URL: https://www.anthropic.com/research/project-fetch-phase-two
  • PDF: N/A
  • DOI: N/A

Plain-Language Brief

A follow-up to "Project Fetch" (August 2025) in which Anthropic's Frontier Red Team ran Claude Opus 4.7 autonomously on the same robotics tasks previously assigned to human-AI hybrid teams. The model completed those tasks roughly 20x faster than the fastest human team from Phase One, while writing far fewer lines of code. The paper also documents where the model failed and attributes the capability gains to general model scaling rather than targeted robotics work.

What the Document Says

  1. Claude Opus 4.7 was run autonomously (three trials, adaptive thinking, maximum effort setting in Claude Code) on the same task set from Project Fetch Phase One.
  2. Minimal human involvement: a researcher plugged in a laptop, entered an initial prompt, and approved commands and task transitions.
  3. Tasks: connect to video camera, connect to lidar sensor, write manual control programs, monitor robot path, detect a beach ball, retrieve the ball autonomously.
  4. On tasks completed by all teams, Opus 4.7 was 37.7x faster than the team without Claude access.
  5. Opus 4.7 generated 1,045 lines of code versus Team Claude's 10,309 lines in Phase One.
  6. The model failed to autonomously complete the final "fetch" step (precise ball retrieval).
  7. The model could not develop low-level robotic control policies or actuation strategies.
  8. Initial algorithm selection was suboptimal (outdated object detection methods used first).
  9. Required human supervision for closed-loop feedback adjustments humans perform naturally.
  10. Authors state: "this progress is not the result of a concerted effort to improve the robotics capabilities" — capability gains are attributed to general model scaling.
  11. Authors describe a trajectory pattern: models help humans → humans help models → models act largely independently. They note this pattern previously appeared in cybersecurity and now appears "at the intersection of AI and the physical world."
  12. Authors describe the current moment as "early era of physical agentic AI capable of utilizing off-the-shelf tools."

Key Concepts Mentioned

  • Physical agentic AI
  • Autonomous robotics
  • Claude Code with adaptive thinking / maximum effort setting
  • Claude Opus 4.7
  • Phase One vs Phase Two comparison
  • Emergent capability from general scaling
  • Lidar, camera, ball detection, manual control programs
  • Human-in-the-loop vs fully autonomous operation
  • Off-the-shelf robotics tooling

Claims Made by the Document

  1. Opus 4.7 (autonomous) completed tasks ~20x faster than the fastest human team from Phase One.
  2. On tasks all teams completed, speedup was 37.7x over the no-Claude team.
  3. Code volume was ~10x lower (1,045 vs 10,309 lines) at higher task performance.
  4. The capability improvement was not the result of targeted robotics fine-tuning.
  5. A repeating pattern of AI capability development ("helpful to humans" → "humans helpful to models" → "models act independently") is manifesting in physical-world robotics.

Evidence Used

  • Direct comparison against Phase One elapsed-time measurements per objective.
  • Qualitative success assessment using the same task definitions as Phase One.
  • Three repeated trials of Opus 4.7.
  • Line-count comparison between Phase One Team Claude output and Phase Two Opus 4.7 output.

Methods / Approach

  • Same task suite as Project Fetch Phase One (August 2025).
  • Three trials, Claude Code, adaptive thinking mode, maximum effort.
  • Elapsed time per objective as the primary metric.
  • Human involvement limited to physical setup, initial prompt, and command/transition approvals.

Limitations Stated by the Authors

  • Model failed to complete autonomous ball retrieval ("fetching").
  • Cannot develop low-level robotic control policies or actuation strategies.
  • Initial algorithm selection was suboptimal.
  • Required human supervision for closed-loop physical feedback.
  • Substantial barriers likely remain for customized hardware and generalized physical adaptation.

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • This is a first-party Anthropic publication. It reports on Anthropic's own model. Independent replication has not been cited.
  • "20x faster" and "37.7x faster" figures compare against Phase One human teams, not against other state-of-the-art robotic AI systems.
  • Task scope (one robot, one controlled environment, defined objectives) is narrow. Generalizability to diverse physical environments is not demonstrated.
  • The document does not compare Opus 4.7 against competing models on the same task suite.
  • "Off-the-shelf tools" framing may understate the degree of researcher scaffolding (laptop setup, prompt entry, approval steps).
  • Source kind is industry blog / first-party vendor content, not peer-reviewed. Claims should be weighted accordingly.

Open Questions

  • How does Opus 4.7 autonomous performance compare to other frontier models on the same task suite?
  • What specifically constitutes the "closed-loop feedback" that still required human supervision?
  • Will Phase Three test generalization to diverse physical environments or different robot hardware?
  • Is the "20x speedup" metric sensitive to how human team timings were measured in Phase One?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • [Fact] The document explicitly states capability gains are from general scaling, not robotics-specific training. This is a factual claim from the authors, not agent inference.
  • [Interpretation] The "helpful to humans → humans helpful to models → models act independently" framing is a narrative arc the authors impose on the data. Whether it constitutes a general law or a post-hoc pattern is not established.
  • Relevant to vault topics: agentic AI, physical AI, capability evaluation, frontier model behavior in open-ended environments.
  • Cross-reference with Phase One (Project Fetch, August 2025) when that note is collected.

Raw Abstract / Excerpt

"Claude Opus 4.7—operating without human assistance—was about 20 times faster than the fastest human team... this progress is not the result of a concerted effort to improve the robotics capabilities."

"We see this trajectory repeatedly: first, models are helpful to humans. Then, humans are helpful to models. Finally, models are largely able to do things themselves."