Anthropic Engineers Explain Claude's Dynamic Workflows at AI Engineer Conference: Letting the Model Orchestrate Multi-Agents
This resource recommendation covers a session from the AI Engineer New York 2026 conference. Anthropic Members of Technical Staff CJ Avilla and Harrison Stall introduced the new 'dynamic workflows' feature, demonstrating how Claude autonomously determines agent topology for complex tasks and discussing engineering boundaries regarding token budgets and sub-agent communication.
At the AI Engineer New York conference that kicked off on October 12, 2026, Anthropic held a workshop on orchestrating its Claude Managed Agents. According to reporting by tech media outlet Latent Space, CJ Avilla and Harrison Stall, both Members of Technical Staff at Anthropic, focused on introducing a new feature called 'dynamic workflows' (or simply 'flows'). This capability aims to solve the bottleneck developers face when managing large numbers of parallel agents, allowing Claude to write scripts itself to coordinate dozens or even hundreds of sub-agents.
The presenters drew lessons from the evolution of Unix and Linux operating systems, noting that as the number of agents grows, the system requires higher levels of organization. In traditional Claude Managed Agents, developers had to manually configure a coordinator and a roster of sub-agents; in dynamic workflows, the lead agent creates a plan, and Claude generates an orchestration program that launches agents, collects results, and manages information flow. Stall stated that this allows users to accomplish discrete, verifiable units of work without having to manage any of the interconnections between them, enabling the handling of projects far larger than a single conversation context.
Beyond the core mechanism, the live Q&A touched on several key engineering challenges in multi-agent systems. When asked how to prevent simple tasks from triggering too many sub-agents and causing cost overruns, Stall replied that it relies primarily on evaluations ('evals') and lower defaults. Regarding token budgets, he revealed that the model does not know its budget while running, but developers can set a session budget. He humorously added that informing models of their budgets could make the agents 'a little bit anxious.' Additionally, Avilla mentioned a research preview feature called 'dreaming,' which helps repeated workflows become more efficient over time by reorganizing memory stores.