--- title: Anthropic says your subagents are just Conway's law description: Anthropic's new Claude Code guide argues that subagent setups often just mirror a team's org chart, a case of Conway's law. Its fix is progressive disclosure:... canonical: https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Anthropic says your subagents are just Conway's law | daily.dev og:description: Anthropic's new Claude Code guide argues that subagent setups often just mirror a team's org chart, a case of Conway's law. Its fix is progressive disclosure:... og:url: https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf og:image: https://api.daily.dev/og/posts/ofrl3bWZF.png og:image:alt: Anthropic says your subagents are just Conway's law og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Anthropic says your subagents are just Conway's law **[Trends](https://daily.dev/sources/trends)** · 3 min read · 0 upvotes · 0 comments ## Summary Anthropic's new Claude Code guide argues that subagent setups often just mirror a team's org chart, a case of Conway's law. Its fix is progressive disclosure: load skills and tools only when needed, which Anthropic says cut the system prompt for advanced models by over 80%. One customer's single agent loop ran about twice as fast as the multi-agent setup it replaced. Tool-choice accuracy also degrades past roughly 30 to 50 tool definitions. @omarsar0 mostly agrees newer models handle large contexts better, but still uses subagents for parallel research and warns to run cost and efficiency evals before removing them. Claude Code creator Boris Cherny has also reversed earlier advice, now favoring Auto Mode and minimal system prompts over heavy context. ## Content Anthropic just published a guide that amounts to a gentle roast of how most of us build agents. Its argument: teams overuse sub-agents by copying their org chart into agent design, which it calls Conway's law at work. The recommendation is one main loop with progressive disclosure, loading skills and tools only when needed. The numbers are the pitch. Anthropic says it cut Claude Code's system prompt for advanced models by over 80%. In one customer case a single agent loop ran roughly twice as fast as the multi-agent setup it replaced. Tool choice degrades past roughly 30 to 50 tool definitions, and the advice is to test changes with as few as 10 golden examples before collapsing sub-agents into one loop. ## The reaction: yes, but run evals @omarsar0 called it a "banger guide" and mostly agrees. Newer models are "more persistent and better at managing large contexts," so subagents "might be getting in the way now." But he won't let anyone just rip them out: "You need to run evals (cost, efficiency, etc.)" first. He still uses subagents for complex research, where parallel context windows earn their keep, and says too many tool calls, context bloat and muddy separation of concerns are still real problems they solve. The how-to crowd isn't backing off either. Walkthroughs of custom subagents (Markdown files in `.claude/agents/`, up to 20 running concurrently by default) sit next to writeups where subagents are the savings: one author says they halved token use by keeping the root CLAUDE.md lean, turning repeated workflows into skills, and sending Docker-log and database-error digging to a subagent that returns only a summary. Another builds whole epics as waves of parallel subagents, one pull request per wave, with a human checkpoint between waves. Boris Cherny, Claude Code's creator, has also reversed some of his own advice. He now skips Plan Mode for synchronous CLI work in favor of Auto Mode, and prefers minimal system prompts and tools over heavy context. ## Meanwhile, a token leak The less flattering story: Claude Code re-sends each active background agent's whole transcript about every 30 seconds to write a 3-5 word status label. One writeup says that eats 15-20% of budget, with telemetry showing agent_summary costing more than the main conversation at roughly 9,000 input tokens per output token. @CobaltHere cites one team measuring 16.4% of spend. Claude Code's team puts it at about 2% on average. There's no official off switch, and it's been on by default since 2.1.232. The community fix is a Bun preload script, no-agent-summary.js, that intercepts setTimeout/setInterval. After applying it, agent_summary dropped to 0% of spend. A guide preaching context discipline, published alongside a harness that burns tokens on a UI label, is a funny pairing. ## Questions this post answers ### Should I use multiple subagents or one main agent loop in Claude Code? Anthropic now recommends defaulting to a single main agent loop rather than multiple specialized subagents, calling heavy subagent use a case of Conway's law where teams copy their org chart into agent design. Using progressive disclosure (loading skills and tools only when needed) cut Claude Code's system prompt for advanced models by over 80%, and one customer's single-agent setup ran roughly twice as fast as its multi-agent predecessor. _Weighing agent architecture trade-offs like this is easier when daily.dev surfaces the latest Claude Code guidance._ ### How many tool definitions can I give a Claude Code agent before performance degrades? Tool-choice accuracy degrades once an agent has roughly 30 to 50 tool definitions, according to Anthropic's Claude Code guidance. This is one reason the guide argues against piling every tool onto a single agent and instead favors progressive disclosure, loading only the tools and skills actually needed for a given task. _Developers tuning agent tool counts can follow evolving Claude Code best practices on daily.dev._ ### How do I create a custom subagent in Claude Code? A custom subagent is a Markdown file with 'name' and 'description' frontmatter placed in '.claude/agents/' or '~/.claude/agents/'. Each subagent runs in its own context window with its own system prompt, tools and permissions, keeping noisy work like codebase search or test output out of the main session. Up to 20 subagents run concurrently by default, configurable via the 'CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS' environment variable. _daily.dev helps developers keep up with Claude Code configuration details like these as they change._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 73 comments across hackernews (as of 2026-10-30). **TL;DR:** Commenters largely confirm the token-burn issue is real but disagree on its actual severity, with some finding the practice indefensible regardless of cost and others suspecting it's tied to profit motives ahead of an IPO. **Sentiment:** 5% positive · 35% mixed · 60% skeptical **The pushback** - Even a small percentage of wasted tokens on background summaries is seen as indefensible when there's no way to turn it off. - Some suspect this pattern of wasteful token burn is a recurring issue with the harness and may be deliberate ahead of an IPO. - Official documentation claims status lines don't consume API tokens, which commenters say contradicts what's being reported. - One person called it fraud, arguing customers are unknowingly paying for a wasteful background process. **By community** - hackernews (mixed): Discussion splits between confirming and quantifying the real cost (with wide variance in reported percentages), debating caching mechanics, and broader griping about token burn, pricing, and trust in the harness provider. **Hottest debate:** Whether the extra cost is actually significant once prompt caching is accounted for, with estimates ranging from near-zero impact to a full 15-20% hit. **Open questions** - Does the issue affect Claude Desktop or the VS Code extension, or only the terminal harness? - Does disabling the agent view setting actually stop the summary calls or just hide the UI panel? - Is the fix Anthropic reportedly rolled out actually the same 'short' status mode mentioned in the original report? **Highlights** > Claude Code's (CC) status summaries used 16.4% of our token spend in a week! We have investigated a source of overhead in CC: it's an agent_summary query that calls your current session/subagent model approximately every 30 seconds for a few words in CC’s status line. Our DRIM system flagged the anomaly in CC's OTEL traces. Grouping requests by query_source showed that over one measured week agent_summary accounted for: 16.4% of total estimated spend (versus 15.7% for the main conversation), and 19.7% of all API requests with approximately 9,000 input tokens per output token. Most requests reuse the prompt cache, but repeatedly reading a large cached transcript still has a cost. Longer contexts and more concurrent agents increase the overhead. We couldn’t find a supported setting that disables only these summaries while preserving background agents and forks. Disabling those capabilities also stops the summaries, but changes the workflow we want to keep. Our workaround is a small Bun preload that intercepts the summary timer without modifying the CC binary. It identifies the timer callback by source-text anchors and leaves other timers alone. In controlled tests on Linux (from ver 2.1.257 to 2.1.312), the same workload produced five summary calls without the hook and zero with suppression enabled. On our current production setup, agent_summary calls remained at zero with the hook enabled. Background agents and forks continued running. The only visible change was that their generated status summaries stopped updating. The workaround depends on implementation details. An update can break the callback matching and silently restore the requests. macOS and Windows are untested. If someone could help us with testing there, it would be much appreciated. If you want to try it - the script, full technical explanation, test results, and installation instructions are on GitHub: https://gist.github.com/flamerix-ai/c3905705cc4205e5de3011fa... > — [NikZaal on hackernews](https://news.ycombinator.com/item?id=50284256) > I tested it, and it looks real but more nuanced. It is just for subagents; workflows do not seem affected. I guess many new users prefer workflows (ultracode), so the impact is less than the report suggested. Moreover it is mostly cached reads, so subagents with short context cost little, while long-context ones cost much more. In my Claude Code sessions, it looks more like ~8% at most. > — [jakozaur on hackernews · 1 comments](https://news.ycombinator.com/item?id=50284683) > Background agents only have 5 minutes' cache TTL, but I'm guessing this doesn't help extend it (not that it would be too useful anyway, especially considering the extra burn). Also, the claim "15-20% of your tokens" seems unclear; of all tokens, or of tokens spent on background agents? Poorly edited AI writing strikes again! > — [creamyhorror on hackernews](https://news.ycombinator.com/item?id=50284641) > Perhaps this is terminology confusion, but the official docs explicitly state that status lines "[Run] locally and [do] not consume API tokens." [0] _If_ that documentation is talking about the same thing that TFA is talking about, then I'd say that this is very much not common knowledge. I also wonder if the fix that they "just rolled out" is actually activating the 'short' mode that TFA mentions... [0] > — [simoncion on hackernews](https://news.ycombinator.com/item?id=50285189) > This is fraud. Unknowing to its customers a company is running a daemon that wastes a commodity it sells to said customer. > — [emsign on hackernews](https://news.ycombinator.com/item?id=50285181) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50284255) · 201 points · 73 comments ## Similar posts on daily.dev - [I set up Claude Code the way Anthropic now recommends, and the sub-agents-spawning-sub-agents trick changed how I work](https://daily.dev/posts/i-set-up-claude-code-the-way-anthropic-now-recommends-and-the-sub-agents-spawning-sub-agents-trick--kltsswmr1) · XDA Developers · 0 upvotes · 0 comments - [I Read the Claude Code Source Analysis So You Don’t Have To](https://daily.dev/posts/i-read-the-claude-code-source-analysis-so-you-don-t-have-to-bvxh2nsri) · InfoSec Write-ups · 0 upvotes · 0 comments - [January 2026: AI Agents Take Over, Claude Code Workflows, Multi-Agent Orchestration, and OpenCode](https://daily.dev/posts/january-2026-ai-agents-take-over-claude-code-workflows-multi-agent-orchestration-and-opencode-g2tclquzi) · Code with Andrea · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#anthropic](https://daily.dev/tags/anthropic), [#claude-code](https://daily.dev/tags/claude-code), [#prompt-engineering](https://daily.dev/tags/prompt-engineering) [View this post on daily.dev](https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Anthropic says your subagents are just Conway's law","url":"https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf"},"datePublished":"2026-10-28T09:24:44.039Z","dateModified":"2026-10-30T17:35:57.934Z","description":"Anthropic's new Claude Code guide argues that subagent setups often just mirror a team's org chart, a case of Conway's law. Its fix is progressive disclosure:...","image":"https://i.ytimg.com/vi/fdtkh30oFA8/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/fdtkh30oFA8/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,anthropic,claude-code,prompt-engineering","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Anthropic says your subagents are just Conway's law"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/anthropic-says-your-subagents-are-just-conway-s-law-ofrl3bwzf#faq","mainEntity":[{"@type":"Question","name":"Should I use multiple subagents or one main agent loop in Claude Code?","acceptedAnswer":{"@type":"Answer","text":"Anthropic now recommends defaulting to a single main agent loop rather than multiple specialized subagents, calling heavy subagent use a case of Conway's law where teams copy their org chart into agent design. Using progressive disclosure (loading skills and tools only when needed) cut Claude Code's system prompt for advanced models by over 80%, and one customer's single-agent setup ran roughly twice as fast as its multi-agent predecessor. Weighing agent architecture trade-offs like this is easier when daily.dev surfaces the latest Claude Code guidance."}},{"@type":"Question","name":"How many tool definitions can I give a Claude Code agent before performance degrades?","acceptedAnswer":{"@type":"Answer","text":"Tool-choice accuracy degrades once an agent has roughly 30 to 50 tool definitions, according to Anthropic's Claude Code guidance. This is one reason the guide argues against piling every tool onto a single agent and instead favors progressive disclosure, loading only the tools and skills actually needed for a given task. Developers tuning agent tool counts can follow evolving Claude Code best practices on daily.dev."}},{"@type":"Question","name":"How do I create a custom subagent in Claude Code?","acceptedAnswer":{"@type":"Answer","text":"A custom subagent is a Markdown file with 'name' and 'description' frontmatter placed in '.claude/agents/' or '~/.claude/agents/'. Each subagent runs in its own context window with its own system prompt, tools and permissions, keeping noisy work like codebase search or test output out of the main session. Up to 20 subagents run concurrently by default, configurable via the 'CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS' environment variable. daily.dev helps developers keep up with Claude Code configuration details like these as they change."}}]} ```