How Claude Changed the Work of Anthropic Researchers: Agents, Parallel Experiments, and the Limits of Performance Evaluation

Anthropic researchers’ work with Claude has gradually shifted from generating individual code fragments to agentic execution of parts of experiments. Company employees describe a system that writes and fixes code, prepares infrastructure, runs computations, and helps compare hypotheses. Choosing an important scientific problem and validating the output remain human responsibilities for now.

The main question here is not whether productivity has increased by exactly two, three, or five times. More important is understanding which parts of the research cycle have already been automated and why one specialist’s subjective speedup cannot be directly extrapolated to all of Anthropic.

Three stages in the transition from coding assistant to research agent

According to interviews with Anthropic employees, the workflow developed in three stages. Before the arrival of Opus 4.5, models more often helped with small fragments of programs under constant supervision. After the release of Opus 4.5, Claude began carrying out useful tasks without continuous observation. With the emergence of the version with Preview and internal agent tools, the system began conducting individual experiments independently within a predefined objective.

This sequence should not be conflated with the official history of all Anthropic products. In the company’s available materials, Opus 4.5 is described as a model for programming, agents, and computer use, while later publications separately discuss long-running tasks, parallel sessions, work delegation, and code execution. Names such as Claude 5, Model 4, or Mythos Preview refer to internal or unconfirmed designations in some of the cited accounts. They cannot be treated as one unified public product line.

What changed in the workflow itself

Previously, a researcher would formulate a task, receive code from the model, transfer it to an execution environment, fix errors, and only then run the experiment. Now an agent can handle most of this chain on its own: examine the repository, propose a plan, modify files, install dependencies, run tests, analyze a failure, and try again.

The difference is not limited to the number of lines generated. Researchers move from a hypothesis to a verifiable result more quickly, while manual work shifts toward defining the task, choosing metrics, evaluating data quality, and interpreting observations.

How agents’ parallel work is organized

Anthropic employees have described a shift from sequential dialogue with a single assistant to running several agents simultaneously. Each agent is assigned a separate task: one modifies training code, another checks an alternative configuration, a third analyzes the results of a previous run, and a fourth handles infrastructure debugging.

This approach works only when the tasks can genuinely be divided. Independent hypotheses, reading different parts of the codebase, and repeatable tests are well suited to parallel execution. Tasks involving shared files, a single experiment state, or sequential dependencies require coordination. Otherwise, agents may overwrite changes, duplicate work, or compare incompatible results.

• Lead researcher formulates the question, sets success criteria, and decides which results merit further work.
• Planning agent breaks the work into independent stages and records the order of actions.
• Executor writes code, runs commands, and saves results in an agreed format.
• Review agent looks for implementation errors, analyzes tests, and compares the conclusions with the original hypothesis.
• Human reviewer checks significant changes, computational expenditure, and the correctness of the scientific interpretation.

Anthropic’s internal tools complement this arrangement with the ability to run several local and remote sessions in parallel. The company’s public materials also describe team modes in Claude Code, where agents distribute independent subtasks and coordinate the results.

This feature does not turn the system into an autonomous laboratory. It increases execution throughput but does not eliminate the need to choose a meaningful question.

Which tasks Claude is already taking over from researchers

The most noticeable area of automation involves programming. Claude creates new files, rewrites existing modules, fixes tests, looks for the causes of errors, and prepares the environment for execution. Anthropic’s official materials say that the model increasingly works not as a prompt generator but as an agent that executes code and checks the result itself.

The boundary runs between execution and goal-setting. An agent handles an experiment well when the goal, metrics, and success criteria are specified in advance. It is significantly weaker when it must decide what problem is worth investigating in the first place, which data deserve trust, and whether an appealing result is the consequence of an error in the test design.

Why the time from idea to experiment matters more than the speed of writing code

One employee estimated that about 80 percent of Claude’s value comes from shortening the gap between an idea arising and launching a test. This is not a precise measurement of company-wide productivity, but rather the personal perspective of a specialist comparing their own workflow with the previous mode.

If implementing an idea previously took a day or two, an agentic tool makes it possible to assemble a minimal test almost immediately. This changes the economics of selecting hypotheses: the researcher abandons a weak direction more quickly and directs attention to the next option. According to employees, it has become possible to test about ten hypotheses simultaneously, although the actual value of this parallelism depends on available GPUs, queue times, the quality of automated metrics, and the complexity of the analysis.

The speedup comes from several sources. The model reduces manual programming, fixes common errors more quickly, retains technical context, and performs routine operations without researcher involvement. Every new run requires computational resources, and the growing number of results creates a new burden for people who must read logs, check anomalies, and choose the next step.

Why productivity estimates diverge

Interviews contain different estimates of AI’s contribution. One researcher spoke of a 50–80 percent speedup, another attributed the result simultaneously to Claude and hiring new specialists, while yet another considered staff growth the main factor. The most skeptical estimate of AI’s contribution was 15–20 percent. These figures cannot be averaged without a unified methodology because employees may have measured different projects, periods, and types of output.

Another employee compared working with Claude to the work of an entire group of graduate students and research staff and considered it unlikely that researcher productivity had increased by less than threefold. This is evidence of the scale of the subjective effect, but not an independent benchmark. The employee is evaluating their own work, knows that the tool is available, and may fail to account for tasks that became mandatory precisely because more automation was available.

Internal Slack surveys also showed a sharp increase in subjective productivity: approximately 1,2 times for Claude 3.7 Sonnet at the beginning of 2025, 5 times for Claude 5 in June 2026, and approximately 12 times for Model 4 in October. These data are presented as reports of internal estimates, not as the results of a rigorously organized experiment. Participants were not selected according to a single scientific protocol, the responses may have contained exaggerations, and the figures themselves were not verified by comparison with a control group.

Individual speedup does not equal company-wide growth

Researcher productivity and organizational productivity measure different things. One specialist may indeed perform several times more technical actions, but the company will not see the same increase if projects are constrained by a shared cluster, computing queues, code review, security, data management, or managers’ ability to choose the best directions.

This is where the bottleneck principle applies. When Claude speeds up code writing and experiment execution, result validation becomes the constraint. When agents generate more hypotheses, human attention becomes the constraint. When computing becomes available to a wider group of employees, demand for GPUs and storage increases. The overall effect is determined by the slowest part of the chain, not by the fastest agent.

For this reason, lines of code per engineer, the number of experiments run, or the subjective feeling that “I did five times more” cannot be used on their own. Evaluating an organization requires information about result quality, the share of changes accepted, time to a reproducible conclusion, computational cost, the number of rollbacks, and the number of projects that reached a useful result.

How to evaluate the impact of AI agents in a research team

A team implementing Claude or another agent should measure not the volume of generation but the full work cycle. First, it should record typical tasks and baseline metrics, then introduce the agent to a limited set of projects while preserving action logs and review results. Comparisons should use identical criteria; otherwise, a faster launch will appear successful even when output quality has declined.

1. Describe the task before connecting the agent: the goal, input data, success criterion, and permissible computational budget.
2. Divide the work into stages: hypothesis formulation, coding, execution, analysis, and decision-making.
3. Record which actions the agent performs independently and where human confirmation is required.
4. Collect not only the time to the first result but also the time to a reproducible result.
5. Have an independent reviewer check the code, data, statistical conclusions, and reasons for abandoning a hypothesis.
6. Compare a series of tasks rather than one successful example, and separately account for the cost of tokens, computation, and human oversight.

But what happens if an agent quickly produces a beautiful chart while using the wrong metric? The answer is simple: speed does not create scientific value without checking the formulation and the result. In research, automation should reduce routine work, not lower the standard of evidence.

The main conclusion: Claude speeds up execution but does not eliminate research judgment

Anthropic materials and employee accounts show a noticeable shift from a coding assistant to an agent capable of carrying out individual stages of a research experiment. The best-supported part of the picture concerns programming, debugging, infrastructure, tools for long-running tasks, and the parallel execution of independent workstreams.

Stronger claims about multiplying productivity, fully autonomous research, and internal models require careful wording. They are based on interviews, internal surveys, company documents, and researchers’ accounts, while some of the material is limited in transparency. An estimate of two, three, five, or twelve times may describe different levels: individual speed, the volume of technical work, the number of tested hypotheses, or the output of the organization as a whole.

The most reasonable interpretation is this: Claude is already reducing the cost of carrying out many well-designed experiments, while the human role is shifting toward selecting problems, designing tests, overseeing resources, and assessing the reliability of the conclusions. If computing power and model quality continue to improve, the main constraint on research teams will not be the ability to write yet another script, but the ability to distinguish an important result from a quickly obtained but useless signal.