On October 22, 2026, Anthropic published a technical guide on gradually consolidating the multiple AI agents that companies have built for individual uses and departments into a single agent loop. In the conventional setup, subagents with different roles each keep their own conversation history and instructions, and take over work handed off by a primary agent. The new guidance recommends revisiting this division of labor as needed, moving to a configuration in which a single AI model repeatedly makes decisions, runs tools, checks results, and revises its plan.
Anthropic says that in multiple enterprise deployments, the single-loop configuration often improved answer quality while also reducing cost and processing time. In one customer comparison, processing time was reportedly about halved compared with the earlier multi-agent configuration. However, the customer's name and the detailed conditions of the comparison have not been disclosed, and the result does not show that the same effect will appear in every system. The guide is technical guidance based on deployment experience, not a peer-reviewed research paper.
Fewer subagents, with decisions consolidated in a single loop
In a conventional multi-agent setup, a primary agent assigns tasks and delegates planning, search, and specialized processing to separate subagents. But because each agent works in its own independent context, information is summarized during handoffs, and important details can be lost. The more agents involved, the more complicated these handoffs become, and processing time and cost tend to rise as well.
In the single loop Anthropic proposes, one model understands the user's request, calls the necessary tools, checks the results, and decides what to do next. Decisions that were previously spread across a router, a planner, and domain-specific subagents are gathered into the same context. When new information turns up mid-task, the plan can be revised in light of it, reducing the need to pass explanations between agents.
That said, not every procedure and tool specification is packed into the system prompt from the start. Anthropic emphasizes "Progressive Disclosure," in which necessary information is loaded in stages. The model is first shown short descriptions of skills and tools, and detailed procedures and definitions are loaded only when they are actually needed. A skill bundles the procedures and decision criteria for a particular task, and serves in place of some of the specialized instructions that subagents used to hold.
This configuration can still run multiple processes in parallel. Tool calls that do not depend on one another are handled simultaneously by the execution infrastructure, and the results are returned to the same model. The model reviews the information in a shared context and decides its next action. In other words, consolidating the agent that makes decisions into one is a separate matter from running processes sequentially.
Anthropic has also disclosed that, as part of design changes suited to higher-performance models, it cut the Claude Code system prompt by more than 80%. However, this is the reduction in the instruction text initially given to the model, not an 80% drop in total token consumption or usage fees. A single loop may reduce cost and processing time, but the effect varies with the nature of the work, the model used, and the tool configuration.
The test for keeping subagents is context, not department
Anthropic does not argue that all subagents should be eliminated. Keeping subagents is more appropriate when the amount of information handled in a single loop grows so large that performance degrades, or when work independent of other tasks is to proceed in parallel. Separating roles is also justified when a single model has trouble choosing the right tool, or when instructions for different tasks conflict. For work requiring independent verification or separation of permissions, subagents remain a valid option.
What matters is not to mirror a company's organizational chart directly in the agent configuration. In its design guide for commerce AI agents, published September 2, Anthropic uses return processing as an example of the problems that arise when work is split up. Deciding whether a return is allowed requires consulting not only order history but also the product catalog and cart contents. If this information is scattered across separate agents, the same customer information may be fetched repeatedly, or the context needed for the decision may be missing.
On the other hand, consolidating the information needed for decisions does not mean centralizing the authority the AI can exercise. In Anthropic's commerce design, even when the AI prepares a purchase, the user confirms the actual order by pressing a button. The APIs the AI calls include no operation that can charge directly, keeping the authority to execute payment separate. Price changes for merchants are likewise first presented as change proposals carrying server-issued IDs, and are applied only after approval by a person in charge; at that point the system rechecks constraints against the conditions at the time of approval.
The October guide maintains this principle. Limits on the number of model runs and on cost, and the conditions for permitting write operations, are managed by the execution infrastructure (the harness) that runs the model, while permission checks and approvals are enforced on the tool side. Skills tell the model how to carry out work, but the final decision on whether something may be executed rests with the code. Sharing the information needed for decisions while restricting authority over critical operations at the execution layer remains an important division of roles in a single-loop design.
On organizational operations, Anthropic recommends a structure in which a small central team manages the system prompt and quality evaluation of the overall workflow, while each department owns skills and tools in its area of expertise. It also notes a tendency for models to have difficulty choosing the right tool once the tool definitions in the system prompt exceed roughly 30 to 50. This is not a strict limit but a rough guide for considering a design that searches for needed tools and loads their definitions at the time of use.
Faster processing is not enough; task completion quality must be evaluated too
Consolidating subagents may shorten processing time by reducing handoffs between models. But Anthropic points out that a design improvement should not be judged by faster processing alone. In a sample Anthropic released, consolidating subagents led to a case where a confirmation step with the shipping specialist was skipped, and a task that had previously succeeded failed.
Specifically, once the primary agent could use the order-lookup tool directly, it generated an answer itself without checking the delivery date with the shipping subagent. Processing time fell, but the necessary check dropped out, and the task failed in one of two attempts. This does not show that a multi-agent configuration is always better; it is an example of how consolidation can cause existing workflow steps to be lost.
To prevent such problems, Anthropic recommends building evaluation tests that verify the completion conditions of the whole workflow. A central team manages shared evaluations covering end-to-end tasks, and each department handles evaluations of the skills and tools it owns. The method is to start evaluating with about 10 correct-answer examples, then integrate subagents one at a time while comparing results before and after each change. However, 10 is a starting point, and that number alone cannot adequately guarantee quality.
Evaluation must look not only at average scores but at how results changed for individual tasks. If a task that succeeded before consolidation begins to fail, it is necessary to investigate whether the cause is tool selection, missing information, or a skipped confirmation step. For workflows that previously used several specialist agents in sequence, it is especially important to test whether the whole chain still completes correctly after consolidation.
The first verification of the migration kit Anthropic released also revealed gaps in evaluation coverage. The initial evaluation did not include tests checking conditions that prohibit returns, or tests covering the end-to-end path from account-related tasks to return processing. Evaluation items were added later, but the results of the first run did not verify the added conditions. That individual functions work correctly and that processes spanning multiple tasks complete all the way through must be evaluated separately.
Real-model verification of the official migration kit is limited
Alongside the guide, Anthropic published on GitHub "single-agent-loop-kit," a reference implementation for migrating an existing multi-agent configuration to a single loop. According to the official README, the kit starts from evaluating the existing system, then proceeds in stages: merging closely coordinated subagents and replacing them with skills and tools. The design runs the same evaluation before and after each change to confirm that quality is maintained through consolidation.
However, not all of the kit's functions were verified using actual AI models. According to the public record, what was tried on a real model was limited to two runs of a sample system, conducted by one person using one type of model. Neither the completed single loop as a whole nor its application to other enterprise systems has been demonstrated.
| Item verified | Scope confirmed in the public record | Unverified / needs confirmation |
|---|---|---|
| Real-model runs | Two runs of a sample system by one person with one model | Reproducibility with other models and in real operations |
| First run on Oct. 8 | 12 tasks evaluated twice per configuration with claude-opus-5 |
Return-prohibition conditions, cross-workflow paths, results on added tasks |
| Whole migration kit | Some merging steps and the approach of calling subagents as tools were run | The completed single_loop, the merging and maintenance stages, migration from a state with no evaluation tests |
| Post-run code changes | Confirmed with offline tests | Quality when the revised configuration is run on a real model |
According to the first run record, the October 8, 2026 test ran 12 tasks twice for each configuration. Tasks were later added to detect failures in cross-workflow processing and existing functionality, bringing the evaluation items to 19. However, the added tasks were not part of the first test. Nor has the comparison between using and not using skills been completed, so the published results alone do not support concluding that replacing subagents with skills is always advantageous.
The official README explicitly lists the scope not verified on a real model, including the completed single loop, the merging and maintenance stages for subagents, migration from systems without evaluation tests, and application to systems other than the sample. For parts where safeguards and evaluation code were improved after the run, confirmation is limited to offline tests. The migration kit is also published as a reference implementation that does not accept maintenance or outside contributions, so companies that adopt it in real operations must take on ongoing management and verification themselves.
Anthropic's new guidance proposes moving from designs that multiply agents and divide functions among them to designs that share the necessary context while supplementing expertise with skills and tools. A single loop is not always optimal, however, and subagents remain an option for independent processing and verification. When considering migration, companies should confirm not only reductions in processing time and cost but also that the completion rate of the overall workflow and the required approval steps are preserved, and choose a configuration based on measured results.