Anthropic: Fold Subagents Into One Loop With Skills for Higher Quality, ~2x Faster
AGI Hunt · wechat · 2026-10-23
Anthropic technical team member Diane Sarkis published a best-practices article on building agents on the Claude Platform; AGI Hunt summarizes it in full. The core finding: a single agent with Skills often beats both a monolithic prompt and a multi-subagent design on quality, usually at lower cost and latency.
Common problems
- Large orgs let each team build its own subagent with a platform team layering an orchestrator on top, so the agent architecture mirrors the org chart (Conway's Law)
- Context loses detail in handoffs; subagents proliferate and overlap
- System prompts become a "Frankenprompt" with conflicting rules and inconsistent tone
- Local evals over-optimize while nobody owns overall quality
- One customer's single loop ran about 2x faster than its multi-agent harness
The alternative
- One model, one context window, with Skills and tool search progressively disclosing context and tools
- Claude Code's system prompt was cut by over 80% for newer models this way
- Since Opus 4.5, a single loop can manage far more than most teams expect
When subagents still make sense
- Measurable degradation from context bloat
- Independent tasks that need parallelism
- A single agent struggles to pick tools or the system prompt has conflicting instructions
Split by context, not by problem type, or agents play a game of telephone.
Migration path
- Start with end-to-end task-completion evals; 10 golden examples suffice
- Merge subagents that are frequently called back-to-back
- Turn retained subagents into tools the main agent calls
- Convert subagent prompts into Skills and API calls into MCP tools, comparing evals one at a time
- Teams maintain Skills, tools and knowledge bases; a small central team owns the system prompt and end-to-end quality
The article notes that how agents get split often depends on who owns what in the company: half architecture, half org design. A migration toolkit and four /claude-api skills are linked at the end.
More from AGI Musings
- The real AI-for-Science question: when experiments become programmable — CatAstro_Piyush · 2026-10-24
- AI scholar Dan Faggella: my ideas will be worthless within 2 years — danfaggella · 2026-10-24
- AGI will force radical human transformation, argues Daniel Faggella in new essay — danfaggella · 2026-10-24
- Building AGI when most humans aren't general: maybe our definition of intelligence needs work — VraserX · 2026-10-24
- Daniel Miessler: code is the lossy output — the primary artifact is the fully articulated ideal in language — JensHonack · 2026-10-24
- Reid Hoffman asks: in 10 years, can an all-AI-agent VC fund beat a human-and-AI firm? — reidhoffman · 2026-10-24