Practitioners on X say small teams with defined roles and explicit handoffs beat single-agent approaches
- OpenAI and Anthropic documentation now make the same practical point practitioners have been trading on X: start with one focused agent, then split work only when roles, tools, or ownership truly differ. - The clearest implementation detail is the handoff contract itself: OpenAI says specialists should take over only when instructions or policy change, while Anthropic packages repeatable know-how in SKILL.md files. - The debate lands as agent builders move from demos to production systems, where reliability depends less on one giant prompt than on explicit workflow design and documented constraints. (developers.openai.com)
The emerging consensus in agent engineering is simple: one giant all-purpose model is not the default answer. Teams are increasingly splitting work across small specialists with explicit handoffs. (developers.openai.com) (anthropic.com) OpenAI’s current Agents documentation says to start with one focused agent, then add specialists only when you need separate ownership, different instructions, different tool surfaces, or different approval policies. (developers.openai.com 1) (developers.openai.com 2) Its orchestration guide draws a hard line between two patterns. In a handoff, a specialist takes over the conversation; in an agents-as-tools setup, a manager stays responsible for the final answer and calls helpers for bounded tasks. (developers.openai.com) That distinction turns “multi-agent” from a buzzword into an operating model. The question is not how many agents you can spawn, but who owns the next decision, what context moves with it, and when control returns. (developers.openai.com) Anthropic describes the same production problem from a different angle. In a June 13, 2025 engineering post, it said its Research feature uses multiple Claude agents, with one agent planning the process and parallel agents searching for information simultaneously. (anthropic.com) Later, on October 16, 2025, Anthropic introduced Agent Skills as folders of instructions, scripts, and resources that agents can discover and load dynamically. Each skill centers on a `SKILL.md` file, with metadata loaded first and the full file pulled in only when relevant. (anthropic.com) That file-based pattern is close to what many practitioners mean when they talk about formalizing handoffs. Instead of hoping one prompt remembers everything, teams write down the job, constraints, and reusable procedures in artifacts another agent can actually load. (anthropic.com) (developers.openai.com) A parallel tool ecosystem is converging on the same structure. GitHub Copilot guidance describes instructions as always-on project context, while skills package reusable multi-step workflows and bundled resources for specialized tasks. (awesome-copilot.github.com) (github.com) Even the caution is shared. OpenAI says “split only when the next branch truly needs different instructions, tools, or policy,” and warns that adding specialists too early creates more prompts, traces, and approval surfaces. (developers.openai.com) So the practical lesson is narrower than the hype. Small teams of agents can beat a single-agent setup when the work has clear ownership boundaries, documented instructions, and explicit handoffs; without that contract, you mostly get extra complexity. (developers.openai.com) (anthropic.com)
