
Anthropic just told teams to delete most of their subagents!
(one customer ran ~2x faster after moving to a single agent loop)
for a year, the default agent design looked like an org chart.
a router on top, a planner, then one specialist subagent per team.
anthropic's applied ai team compared these setups across several enterprise deployments.
a single agent with skills frequently won on quality, and often on cost and latency too.
the problem is handoffs.
every subagent returns a summary, and every summary drops detail. anthropic calls it a telephone game.
here's the loop they recommend instead:
one model, one context window, an open-ended tool loop with turn and cost budgets set in the harness
each turn the model emits a batch of tool calls, the harness runs them in parallel and returns raw results
each subagent's procedure becomes a skill, and its api calls become mcp tools
hard limits like approval gates live inside the tools, not in the prompt
skills and tools load on demand through progressive disclosure and tool search
progressive disclosure is why this works now.
anthropic cut claude code's system prompt for advanced models by over 80%, partly because of it.
there's a catch, though.
on a sample support agent, folding one subagent lowered latency but broke a process that had been passing.
that break showed up in only 1 of 2 runs.
so you can't skip evals here. start with 10 golden examples, fold one subagent at a time, and run every task more than once.
another limit to respect: past 30 to 50 tool definitions, tool choice starts to degrade.
subagents still earn their place for parallel work, for context bloat you can measure, and for tool restrictions.
my take: most multi-agent systems i see were built for weaker models and never revisited.
your agent architecture shouldn't mirror your org chart.
Arindam Majumder 𝕏 @Arindam_1729
I open-sourced Claude Duet!
(same proactive agent, on open models, MIT-licensed)
It's called OpenDuet.
it briefs you every morning, checks your apps while you're away, and only pings you when something needs you.
Anything that sends, posts, or pays waits for your approval.
Each bot gets its own open model and a plain MEMORY.md you can edit.
built with
@mastra,
@nebiustf and
@composio
Repo: