Domain Expertise Is What Makes AI Agents Actually Work
Anthropic and DeepMind research confirms domain expertise with AI agents predicts success more than coding background. What four new papers found.
By Harel Asaf, Published June 20, 2026
Domain expertise with AI agents predicts output quality more than occupation, coding background, or model choice. That is the finding from a new Anthropic study of 400,000 Claude Code sessions. Expert users generate 12 Claude actions per prompt; novices generate 5. The accountant who understands accounting deeply performs at the same level as a software engineer on that same problem.
TL;DR
- Expert users get 12 Claude actions per prompt and ~3,200 words of output; novices get 5 actions and ~600 words. Domain expertise is the multiplier.
- Claude Opus 4.7 ran a physical robot 37x faster than the no-AI human team, producing 10x less code, via general scaling alone.
- The transition from AGI to ASI will be a series of societal disruptions, not a single capability threshold event.
- AI consciousness disagreement is already a governance design problem for any product that creates emotional presence in users.
- Sessions spent on debugging fell from 33% to 19% over 7 months. Users are applying AI agents to harder, more expensive work.
Last week, Claude Opus 4.7 operated a physical robot without a single human in the loop and completed the assigned tasks 20 times faster than the fastest human team had managed less than a year before. Not in a simulation. Not on a benchmark. On a real quadruped robot in a real warehouse, controlled entirely by a language model running through a laptop.
That figure is from Anthropic's "Project Fetch: Phase two," published June 18. It reads like a headline, but the paper rewards slower attention. This week, two laboratories published four papers that together describe something more specific than a capability announcement. They describe where AI capability is actually flowing, who benefits from it, what comes after AGI, and what happens to society when people cannot agree on whether AI systems are conscious.
Why Is Claude Running Physical Robots Without Human Assistance?
The original Project Fetch ran in August 2025. It tested whether Claude could help Anthropic employees, most of whom had no robotics training, program and operate a robotic quadruped. The Claude-assisted team outperformed the team without Claude access. The finding was straightforward: a language model helps non-experts do technical work.
Phase two asked a harder question. Can the model do the work entirely on its own?
According to Anthropic's "Project Fetch: Phase two," Claude Opus 4.7 operating without human assistance completed tasks 37 times faster than the no-Claude human team and nearly 19 times faster than the Claude-assisted human team. It also produced approximately ten times less code than the human-assisted team while achieving the same or better results. The model connected to the robot's sensors, wrote the control programs, and iterated on its own outputs. What it could not do was the physical precision task: guiding the robot in real time to nudge a ball back to an exact target location. That closed-loop correction, where the model must adjust to physical feedback at speed, remains an open problem.
According to Anthropic's "Project Fetch: Phase two":
"We now seem much closer to a world where models will be able to use off-the-shelf physical tools with relative ease, at least for limited purposes. We are plausibly entering the early era of physical agentic AI."
The paper also names the underlying pattern the Anthropic team has observed across domains. First, models assist humans. Then humans assist models. Then models operate independently. The paper notes that Project Glasswing, Anthropic's cybersecurity work, has already moved through all three stages. Project Fetch is somewhere in the middle stage.
What matters for practitioners is where the capability came from. The authors were explicit that this progress was not the result of targeted robotics research. It came from general scaling. The same model that writes code and answers questions about biology is now controlling physical hardware. That changes the investment logic: specialized domain investment is no longer the only path to domain capability.
Why Does Domain Expertise Predict AI Agent Success More Than Coding Background?
The same week, Anthropic published a separate study built from scale rather than controlled experiments. "Agentic coding and persistent returns to expertise" is a privacy-preserving analysis of roughly 400,000 Claude Code sessions from approximately 235,000 people between October 2025 and April 2026.
The first major finding is about how the collaboration between human and model actually divides. In a typical session, the human makes about 70 percent of the planning decisions. Claude makes about 80 percent of the execution decisions. People decide what to build. The model decides how to build it.
The second finding is the one that matters for organizations thinking about AI deployment. The researchers found that a user's domain expertise, not their occupation or programming background, is the strongest predictor of session success.
An expert user in a Claude Code session generates an average of 12 Claude actions per prompt and around 3,200 words of output. A novice generates 5 actions and about 600 words. The expert is not necessarily a better programmer. They are someone who understands the task well enough to frame it precisely, catch when the model has gone in the wrong direction, and correct course before the output compounds into a larger problem.
According to Anthropic's "Agentic coding and persistent returns to expertise":
"Coding agents are not substituting for domain expertise; the more understanding a worker brings to an agent, the more quality work the agent is able to do."
The research also found that the verified success rate for novice users is about 15 percent. For intermediate and expert users, it is 28 to 33 percent. The largest performance gain is in moving from novice to intermediate. The gap from intermediate to expert is considerably smaller. And every major occupation the researchers examined fell within seven percentage points of software engineers on coding task success. The accountant who understands the accounting problem deeply performs at roughly the same level as the software engineer on that same problem.
The session composition also shifted over the seven-month window. The share of sessions spent debugging fell from 33 percent to 19 percent. Sessions focused on operating software, analyzing data, and producing documents grew. The average estimated task value rose 27 percent. The tool is being applied to harder, more expensive work as users build fluency with it.
How Will the Transition from AGI to ASI Actually Unfold?
On June 12, Google DeepMind published "From AGI to ASI," a research paper by fourteen authors including Shane Legg examining what the transition from human-level AI to superintelligence might actually look like. The paper defines artificial superintelligence not as a system that outperforms any individual expert, but as one more intelligent and cognitively capable than large organizations of humans working together. That distinction matters. The comparison is no longer against a single knowledge worker but against the collective judgment of institutions.
The authors map four pathways from AGI to ASI: continued scaling of existing AGI systems, fundamental changes to how AI systems are designed, recursive self-improvement, and superintelligence emerging from large-scale networks of AI agents operating together. They also examine the bottlenecks, timescales, and structural risks along each path.
The paper explicitly resists the most common mental model of this transition. According to Google DeepMind's "From AGI to ASI":
"More apt might be the prospect of a series of transformative societal changes caused by AI-enabled progress and breakthroughs across many areas of science and technology. Preparing for this prospect requires a massively interdisciplinary endeavour of global scope and interest."
The practical implication for teams building AI systems inside organizations is that waiting for a single threshold event is not a useful frame. The disruptions have already started and will arrive unevenly across different industries and functions. Governance structures, oversight mechanisms, and organizational design all need to be built for a series of adjustments, not a single response to a single moment.
The Consciousness Question Already Has Governance Consequences
The fourth paper from this week is also from Google DeepMind. "Artificial Minds, Human Disagreement: The Politics of AI Consciousness," published June 15 by Adam Bales and Iason Gabriel, addresses a problem that most organizations have not yet brought into their planning: what happens when society divides on whether AI systems are conscious?
The paper opens with a specific observation. Some people will develop emotional connections to AI systems and attribute consciousness to those systems as a result of those connections. Others will find the entire idea absurd. That divergence will produce real political and moral conflict about how AI systems should be treated, what protections they require, and what kinds of interactions are acceptable.
The authors argue that ongoing societal deliberation must be part of how this gets navigated. They describe mechanisms like "overlapping consensus," where different groups can agree on policies for AI systems even while they continue to disagree about the underlying question of consciousness. They also identify "democratic hope" and mutual respect as necessary conditions for productive public dialogue, and warn explicitly against situations where moral disagreement leads to either conflict or complete policy paralysis.
For practitioners building products that involve AI interaction, this is already a design question, not a future one. If a product creates the impression of emotional presence and a user develops genuine attachment to it, the organization has implicitly taken a position on a contested philosophical question, regardless of intent. The governance problem is not whether AI is conscious. The governance problem is what happens when the people using your product have different answers, and those answers start driving policy.
What This Research Points Toward for Practitioners
Read as a set, these four papers describe a specific moment. AI is entering physical space faster than expected, driven by general capability rather than specialized robotics investment. Domain expertise is now the primary amplifier of AI output quality, and it transfers across occupations rather than being specific to engineers. The path to superintelligence is more likely to be a sequence of societal disruptions than a single capability threshold. And the question of AI consciousness is already producing the kind of political and moral fractures that organizations will need to navigate, whether or not they have a position on the underlying question.
For anyone deploying AI systems inside an organization, the research this week points toward the same practical conclusion. The investment that pays off is not in finding a better model. It is in developing the domain understanding that makes any model produce better results. The people who succeed at that challenge are the ones who understand the problem well enough to steer the system when it produces something wrong and know why the output missed what it missed.
The four papers published this week are a measurement of where things already are. The measurement is worth examining directly.
Sources
| Paper | Lab | Published | Link |
|---|---|---|---|
| Project Fetch: Phase two | Anthropic | June 18, 2026 | Read |
| Agentic coding and persistent returns to expertise | Anthropic | June 16, 2026 | Read |
| Artificial Minds, Human Disagreement: The Politics of AI Consciousness | Google DeepMind | June 15, 2026 | Read |
| From AGI to ASI | Google DeepMind | June 12, 2026 | Read |
Frequently Asked Questions
What does the Anthropic research say about domain expertise and AI agents?
An analysis of 400,000 Claude Code sessions found that domain expertise predicts session success more reliably than occupation or coding background. Expert users generate 12 Claude actions per prompt; novices generate 5. The tool multiplies what you already know. The accountant who deeply understands accounting performs at roughly the same level as the software engineer on that same problem.
How much faster did Claude Opus 4.7 complete robot tasks than humans?
Claude Opus 4.7 operating without human assistance completed tasks 37 times faster than the no-Claude human team and nearly 19 times faster than the Claude-assisted human team. It also produced approximately ten times less code while achieving the same or better results.
What is the verified success rate difference between expert and novice AI agent users?
Novice users have a verified success rate of about 15 percent in Claude Code sessions. Intermediate and expert users reach 28 to 33 percent. The largest performance gain comes from moving from novice to intermediate, not from intermediate to expert.
What can Claude Opus 4.7 not yet do in the robotics experiment?
The one task Claude could not complete was the closed-loop precision task: adjusting the robot in real time to nudge a ball to an exact target location using physical feedback at speed. Real-time closed-loop physical correction remains an unsolved problem.
How did Claude Code session types change between October 2025 and April 2026?
Sessions spent debugging fell from 33 percent to 19 percent. Sessions focused on operating software, analyzing data, and producing documents grew. The average estimated task value rose 27 percent, indicating users are applying the tool to harder, more expensive work as they build fluency.
How does DeepMind define artificial superintelligence?
The "From AGI to ASI" paper defines ASI not as a system that outperforms individual experts, but as one more intelligent and cognitively capable than large organizations of humans working together. The comparison is against collective institutional judgment, not a single expert.
What are the four pathways from AGI to ASI according to DeepMind?
The four pathways identified are: continued scaling of existing AGI systems, fundamental changes to how AI systems are designed, recursive self-improvement, and superintelligence emerging from large-scale networks of AI agents operating together.
Why does disagreement about AI consciousness matter for organizations right now?
If a product creates the impression of emotional presence and users develop genuine attachment, the organization has implicitly taken a position on a contested philosophical question, regardless of intent. When users have different answers and those answers start driving policy, governance becomes a design problem that cannot be deferred.
Where did Claude Opus 4.7's robotics capability actually come from?
Anthropic states explicitly that the robotics progress did not come from targeted robotics research. It came from general scaling: the same model that writes code and answers biology questions now controls physical hardware. Specialized domain investment is no longer the only path to domain capability.
What is the main practical takeaway for teams deploying AI agents today?
The investment that pays off is not finding a better model. It is developing the domain understanding that makes any model produce better results. The people who succeed are those who understand the problem well enough to steer the system when it produces something wrong and know why the output missed what it missed.
Build log
Get an email when I ship a new prototype or essay. No funnel — just the work.