AI handles everything from planning to assigning tasks
Anthropic has added dynamic workflows to its AI agent execution platform, Claude Managed Agents, allowing multiple agents to divide up work. A lead agent develops a plan, assigns tasks to subagents and consolidates their results. Up to 1,000 agents can run in parallel in a single execution.
An AI agent is an AI system that works toward a given goal. With this new feature, rather than assigning everything to a single agent, tasks are divided between a lead agent and agents that carry them out. The existing execution platform now includes a system for coordinating multiple AIs.
The 1,000-agent limit is not the only point to note. The AI itself handles dividing up the work and consolidating the results. However, the number of agents that can run in parallel does not directly translate into faster processing or proportionally better results.
Detected 66 of 70 bugs
In a test presented by Anthropic, researchers intentionally planted 70 bugs in 116,000 lines of code and compared detection results. A single agent found 14 to 27 bugs in one run, while the dynamic workflow reportedly detected 66 consistently.
That translates to a detection rate of about 20% to 39% for a single agent and about 94% for the dynamic workflow. In this test, dividing work among multiple agents led to more bugs being found. Even so, the result of 66 means that not all of the planted bugs were detected.
These results also came from a specific task: detecting bugs in code. It is not clear whether similar differences would emerge in other types of work, such as research or document writing. The materials provided also do not state how many agents were actually used in the test, so it would be premature to assume the results came from running 1,000 agents.
For adoption, the key measure is cost per result, not the number of agents
Dividing up work also increases the amount of AI processing required. Anthropic warns that the approach may consume a large number of tokens and recommends starting with a small-scale setup. Tokens are units of text or code processed by AI and are a useful measure when considering usage costs.
What companies need to determine is not how many AIs they can run, but how much it costs to achieve the required quality of results. Even if more bugs are found, the decision to adopt the approach will depend on how much costs and verification work increase. A practical approach is to compare a single agent with a small multi-agent setup on the same task, measuring the number of bugs detected, time required and cost.
The materials provided do not give the total cost, processing time or token consumption for this comparative test. The only confirmed difference is in the number of bugs detected; the results do not demonstrate cost-effectiveness.
Comments