Decoupling brain and hands
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A software architecture concept in Claude Managed Agents that separates the brain (the orchestration harness responsible for reasoning and decision-making) from the hands (the execution environments that carry out actions), enabling more reliable and secure long-horizon agent tasks.
Core idea
Lance Martin identifies decoupling brain and hands as a foundational design theme for long-horizon agent systems. The brain—the harness—handles cognition and planning, while the hands—the execution environments—perform concrete actions such as running code, browsing the web, or manipulating files. These two layers are kept structurally separate rather than tightly coupled in a single process. 53:39
Implementation via stateless process and append-only session log
In practice, the decoupling is realized by implementing the harness as a stateless process paired with an append-only session log. Martin argues this specific implementation choice is what makes the architecture both reliable and safe: because all state lives in the log rather than in ephemeral process memory, the harness can crash and resume, be inspected, or be audited without losing or corrupting task state. 1:08:50
Reliability and safety benefits
Martin explicitly advocates for this pattern as appropriate for long-horizon tasks, stating that "this decoupling actually makes it quite reliable and safe, particularly for long-horizon tasks." The separation of concerns limits the blast radius of failures in the execution environment—the hands can fail or be sandboxed without corrupting the reasoning layer—and the stateless-plus-log design allows recovery and auditability. 1:08:50
Points of disagreement
No speakers in the available material contest or critique this concept; it is presented solely from Martin's perspective as a design recommendation within the Claude Managed Agents architecture.