Anthropic's dynamic workflows let Claude orchestrate its own multi-agent topology
Latent Space · rss · 2026-10-13
At AI Engineer NYC, Anthropic engineers CJ Avilla and Harrison Stall demoed "dynamic workflows" (flows), a new Claude Managed Agents feature announced Friday.
- Instead of developers configuring a coordinator and subagent roster, Claude itself plans tasks, writes an orchestration program, launches agents, and manages information flow — all inspectable via Claude Console's graph view.
- Q&A: agent counts are governed by evals and conservative defaults; the model doesn't know its token budget at runtime (developers can set session budgets — telling it might make agents "anxious").
- The "dreaming" research-preview feature reorganizes memory stores (merging duplicates, surfacing insights) to make repeated flows cheaper; subagents share a filesystem but can't message each other via API.
- A stock watchlist scanner demo showed flows cost "much more" than single-agent but find far more detail. Anthropic frames flows as throwing test-time compute at hard problems, aiming to make large-scale scientific work accessible to general users.
More from coding & agent
- Google's RRSI adds regularization to recursive self-improvement, +3.3 points OOD — jiqizhixin · 2026-10-13
- "You have a magic coding robot and your big idea is cloning mediocre apps?" — rms80 · 2026-10-13
- Open-source coding agent Aster joins NVIDIA Inception program — saheedniyi_02 · 2026-10-13
- Solo dev's Bezalel hits $100 MRR: one MCP URL gives agents memory, email, money — Rasmic · 2026-10-13
- How Atomic Agent's desktop app is just a plain HTTP client to its OpenAI-compatible server — GapNew4766 · 2026-10-13
- Author bridges Gemini web chat into a native Linux executor, cutting Antigravity CLI turns from 400+ to 0 — Icy-Management-5843 · 2026-10-13