News · 2026-10-29

Anthropic argues shared work often needs one agent loop

Anthropic says a single agent using progressively loaded skills frequently beat both oversized prompts and subagent arrangements in its enterprise deployments. Diane Sarkis’s October 22 article argues that shared work can suffer when organizations copy their internal departments into separate agents. The finding is a practical design recommendation backed by vendor-reported experience, with explicit reasons to keep useful subagents.

Key facts

• Diane Sarkis published the design article on October 22; it resurfaced in this briefing’s discussion window.
• Anthropic reports one customer achieving roughly twice the runtime speed after consolidation.
• The company says progressive disclosure contributed to reducing Claude Code’s system prompt by “over 80%.”
• Primary source: Anthropic’s agent-design article.

The failure mode is familiar outside artificial intelligence. A customer asks one organization to solve a problem, but the request moves among departments, each carrying a partial account of what happened. The handoffs become work of their own. Anthropic says an agent system can reproduce that friction when every business specialty gets its own model loop and an orchestrator has to translate between them.

A separate agent is not merely a labeled function. It can have a fresh context, its own instructions, a tool set, and several internal calls before returning a summary. Those properties can be useful, but the summary may discard a detail that later proves important. Another agent then has to recover the connection, ask for it again, or proceed without it. The multi-agent systems lesson explains why coordination has a cost even when every participant is competent.

Anthropic’s proposed alternative keeps a main loop responsible for the evolving task. Skills provide procedures and judgment; tools provide actions and hard constraints; retrieval supplies reference facts when needed. Instead of placing every detailed instruction in the opening prompt, the system first exposes short descriptions and retrieves the relevant detail at the moment it becomes useful.

Think of a technician carrying an index of manuals rather than reading every manual before touching the machine. The index preserves discoverability, and the full manual remains available. The saving comes from avoiding irrelevant detail while retaining continuity about the actual repair. In an agent, that can reduce repeated context transfers and let one reasoning process connect evidence across successive steps.

The quoted prompt reduction is a claim about that broader design, not a measured estimate of savings from one isolated change. Anthropic does not provide before-and-after token counts or separate the contribution of every technique. A smaller opening prompt also does not imply that the whole task uses 80% fewer tokens: later skill retrieval and tool work still consume resources. Our context-window lesson helps distinguish what is initially visible from everything processed during a task.

The strongest objection is that one loop can become overloaded. Anthropic agrees. Its post retains subagents when a growing context causes measurable degradation, when independent work benefits from true parallelism and fresh contexts, or when tool choice and conflicting instructions justify separation. Verification is another legitimate reason to introduce a distinct context. The recommendation is to split on actual work boundaries rather than automatically assigning an agent to each box in an organization chart.

There is an unusually useful caveat inside the post itself. A support-agent example became faster after consolidation but broke a delivery-date flow in one of two runs. That does not prove consolidation fails generally. It demonstrates why a clean architecture diagram and lower latency are insufficient evidence that the system still completes the user’s whole job.

Anthropic therefore recommends end-to-end evaluation and incremental migration. Fold in one agent, inspect which outcomes change, and retain separation where it solves a measured problem. Our ablation-studies lesson describes the underlying habit: remove a component and check what it was actually contributing before declaring it unnecessary.

For teams adopting the advice, the comparison should preserve the same tasks and acceptance criteria. Faster responses are useful only if the system still resolves the request correctly. The post’s broken delivery-date example makes that a concrete concern rather than a theoretical warning about simplifying architecture.

The story fits a larger moment in which GitHub’s cloud sessions and OpenAI’s Dots make ongoing agent work a product feature. Persistence does not require a team in every case, and a team does not automatically deliver better continuity. The practical lesson is to measure completed work, latency, and total cost together. Anthropic’s evidence is useful deployment experience, but its unpublished customer details prevent treating the recommendation as a universal experimental law.

Primary source, verified: read the paper →

Ground Truth. (2026, October 29). Anthropic argues shared work often needs one agent loop. Ground Truth. https://groundtruth.day/news/anthropic-single-agent-progressive-disclosure.html

@misc{groundtruth:anthropic-single-agent-progressive-disclosure,
  title  = {Anthropic argues shared work often needs one agent loop},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/anthropic-single-agent-progressive-disclosure.html}
}

Topics: anthropic · agents · developer-tools · architecture · efficiency