OpenHands argues that a single AI agent with broad tools can avoid some problems created by dividing work among specialized agents. Drawing on its experience building software development agents, it points to rigid roles, lost context during handoffs and added maintenance as reasons to question a multi-agent design.
Consider its example of a software development system: a verifier checking a fix might need a file-localization tool assigned to another agent. OpenHands says passing along only a fault localizer’s summary can lose information the next agent needs. Separate prompts or codebases can also make a system larger and more complex. These are OpenHands’ design concerns, not results from a comparative performance test in the post.
One agent still needs instructions
OpenHands describes giving a general-purpose model broad tools and either combining prompts or retrieving instructions as needed. It says its current approach combines prompts, but longer prompts can cost more and excess context can distract models. Choosing how best to supply instructions remains, in its words, “an active research question.” The post still favors multiple agents when one has access to privileged information or different agents act for different people.
Claude’s article on agent design makes a related case for folding some subagents into the main agent loop. The article says that, in comparisons across several enterprise deployments, a single agent with skills frequently outperformed both all-in-one-prompt and subagent designs on quality, often at lower cost and latency per task. It also reports that one customer saw roughly twice the runtime speed with a single-agent loop compared with its multi-agent harness. Those are the Claude article’s reported comparisons, not an OpenHands finding or a universal speed claim.
The Claude article recommends building an end-to-end evaluation, turning a subagent’s prompts into skills and its API calls into tools, then comparing results after each change. In its support-agent example, folding a subagent lowered latency but broke a process that had previously passed. The article still sees a role for subagents in verification, parallel work and modular permissioning, including when a single loop suffers measurable performance degradation from too much context. Its advice is to test task completion before deciding which subagents to keep.
Comments