Anthropic deployment recap reports roughly halved runtime after agent simplification

• Simplifying one customer's multi-agent harness produced roughly 2x faster runtime, according to the deployment recap.
• One main agent can load Skills and tools on demand, while the harness enforces budgets and approvals, as the architecture sketch describes.
• Subagents retain a role in parallel work, heavy context and isolated verification, according to the thread's conclusion.

Anthropic says it cut Claude Code's system prompt by over 80%, partly through progressively loading skills and tools, in its October 22 deployment writeup. The same report describes a support agent that became faster while skipping a required delivery-date lookup.

The speedup came from an unnamed customer comparing its multi-agent harness against a single agent loop. Across several enterprise deployments, Anthropic says a single agent with skills frequently beat both specialist-agent designs and one-prompt-for-everything designs on quality, often with lower cost and latency per task.

Another company found that a dozen specialists made its customer-facing agent's brand voice inconsistent and added latency. An org chart makes a lousy context boundary.

Anthropic's proposed architecture keeps planning and execution in the same model, which revises its plan as results arrive:

• Main loop: One model and one context window, with batches of tool calls executed in parallel and raw results returned to that model.
• Skills: Domain procedures, ordering and judgment calls formerly held in specialist prompts, loaded when needed.
• Tools and MCP: Actions, live data and structured results, with permission checks and hard preconditions enforced inside tools.
• Knowledge bases: Policies, metric definitions and product specifications behind search or lookup tools, or reference files inside skills.
• Harness: Turn and cost budgets, write gates and stopping conditions, with an optional verifier check after the answer.

Tool descriptions also consume context. Anthropic says tool selection tends to degrade beyond 30 to 50 definitions in the system prompt, and describes deferred loading through tool search.

Anthropic describes three ways to reduce a specialist fleet:

1. Merge subagents: Combine specialists that share data, chain on each other's outputs or are frequently called together.
2. Call subagents as tools: Keep the main agent in charge of the conversation while specialists return results. Nested model loops and lossy summaries remain.
3. Replace subagents with skills and tools: Convert prompts into skills and API calls into MCP tools, comparing the folded and separate versions one at a time.

A support agent in Anthropic's migration test got faster after consolidation but answered using an order lookup without consulting the shipping specialist for the delivery date. The broken flow appeared in only one of two runs.

The evaluation setup described in the guide includes:

• Shared end-to-end tests: A centrally owned suite, supplemented by smaller skill-specific evals. Anthropic gives 10 golden examples as a starting point.
• Selection checks: Routing accuracy before consolidation; correct skill loading and tool selection afterward.
• Per-task comparisons: Which production flows changed, alongside aggregate scores, with repeated runs to expose intermittent failures.
• Targeted replay: Timestamps on important events to reconstruct tests where internal systems or user interactions are difficult to replay.

Subagents keep intermediate tool calls and results out of the parent conversation, returning only their final message under the SDK's documented behavior.

Anthropic identifies three conditions that can justify retaining them:

• Measurable performance degradation from context growth in a single loop.
• Independent parallel tasks with fresh context windows that can be discarded afterward.
• Persistent tool-selection problems or conflicting instructions in the main agent.

Loading skills into one session creates its own context-management problem. In dexhorthy's discussion, a five-step workflow seems manageable, while long processes with accumulating instructions or hundreds of active skills may require fresh contexts.

The post distinguishes three kinds of skill content:

• Information: Facts absent from the model's weights, which can often be combined into one message.
• Procedure: Ordered steps whose instructions may need to remain active during later stages.
• Behavior: Steering whose adherence can decay over a long session; the post points to repeated rules or deterministic hooks.

Dexhorthy also described deleting a browser-testing skill after newer models could perform its workflow from the instruction “test in the browser.”

Colin Jarvis, OpenAI's head of Forward Deployed Engineering, reported moving a semiconductor customer's workflows onto a common Codex harness. Custom applications became plugins and skills, while separate terminal UIs and CLIs were retired, as a related recap also noted.

OpenAI says it removed roughly 90% of the custom code from each project. In one August 2026 example, the initial deployment took two months; extending the solution to a second team took one day.