Dynamic Workflow
An AI agent execution pattern in which an agent dynamically determines at runtime which sub-agent or tool to invoke next based on the current state. In contrast to static pipelines with predefined steps, dynamic workflows allow task paths to be adjusted in real time as context changes.
Timeline (last 90 days)
The dynamic workflow approach suits scenarios with fuzzy task boundaries that require exploration, but it also brings reduced predictability: developers cannot know in advance which modules the agent will create or how it will coordinate them.
Boris described dynamic workflow as the 'fourth scaling factor' after model size, data volume, and training compute.
Dynamic workflow is a new feature released by Anthropic a few weeks ago that can dynamically orchestrate dozens, hundreds, or even thousands of sub-agents.
In the Dynamic Workflows demo, the stock watchlist scanner has about 6 concurrent agents at each step, and theoretically supports up to 1000.
In Dynamic Workflows, each sub-agent has an independent context window that is physically isolated at the database level, and by default they do not leak to each other.
Anthropic launched Dynamic Workflows last Friday, in which the main agent formulates a plan and then automatically spawns a network of sub-agents, supporting dynamic increases and decreases in the number of concurrent agents.
Current Dynamic Workflows only support one level of nesting and do not support extremely complex large-scale recursive topologies.
Dynamic Workflows support configuring inference-time compute by stage, for example using Haiku in the scanning stage and Opus in the review stage.
Anthropic positions dynamic workflows as the first step toward directing test-time compute at difficult problems and making scientific discovery broadly achievable by the public.
In dynamic workflows, sub-agents share a file system and can exchange information through it, but developers cannot directly instruct one workflow sub-agent via the API to send a message to another.
14 more timeline events
All Facts (20)
The dynamic workflow approach suits scenarios with fuzzy task boundaries that require exploration, but it also brings reduced predictability: developers cannot know in advance which modules the agent will create or how it will coordinate them.
50%UnverifiedBoris described dynamic workflow as the 'fourth scaling factor' after model size, data volume, and training compute.
50%UnverifiedIn the Dynamic Workflows demo, the stock watchlist scanner has about 6 concurrent agents at each step, and theoretically supports up to 1000.
50%UnverifiedCurrent Dynamic Workflows only support one level of nesting and do not support extremely complex large-scale recursive topologies.
50%UnverifiedDynamic Workflows support configuring inference-time compute by stage, for example using Haiku in the scanning stage and Opus in the review stage.
50%UnverifiedIn Dynamic Workflows, each sub-agent has an independent context window that is physically isolated at the database level, and by default they do not leak to each other.
50%UnverifiedIn dynamic workflows, sub-agents share a file system and can exchange information through it, but developers cannot directly instruct one workflow sub-agent via the API to send a message to another.
50%UnverifiedDynamic workflows are defined as an entirely new way of multi-agent orchestration, built specifically for the most ambitious workloads.
50%UnverifiedDynamic workflows allow Claude to decide and orchestrate its own multi-agent topology, rather than developers pre-statically configuring a coordinator and a roster of sub-agents.
50%UnverifiedUsers can still pass in pre-configured expert sub-agents in the agent configuration of dynamic workflows.
50%UnverifiedThe model does not know at runtime how much available budget (token budget) it has, but developers can set a session budget.
50%UnverifiedDevelopers can always inspect sub-agent behavior through the graphical features within the Claude Console, including viewing the payload the lead agent passes to sub-agents.
50%UnverifiedWhether dynamic workflows let a stock watchlist flow self-improve (such as identifying that the scanning model is insufficient and changing the next plan) is not a built-in feature, but hill climbing and dreaming can be combined to improve the loop.
50%UnverifiedAnthropic positions dynamic workflows as the first step toward directing test-time compute at difficult problems and making scientific discovery broadly achievable by the public.
50%UnverifiedThe orchestration layer of a dynamic workflow is responsible for three things: task decomposition, agent routing and result aggregation.
50%UnverifiedDynamic workflows allow subsequent steps to be adjusted at runtime based on intermediate results, unlike static DAGs.
50%UnverifiedDynamic workflows are better suited to scenarios requiring multi-step reasoning and the coordination of multiple tools, such as refactoring a large codebase or cross-modal content generation.
50%UnverifiedAnthropic says dynamic workflows have been trialed internally in scenarios such as code migration and document generation.
50%UnverifiedThe core idea of Anthropic's dynamic workflow is to let multiple specialized agents each do their own job, coordinated through a workflow engine, rather than scaling up a single model.
50%UnverifiedAnthropic's dynamic workflow is currently closer to a "scheduler" stage, and gaps remain in observability, security, permissions and cross-vendor interoperability before it becomes a complete agent operating system.
50%