Don't Build an Agent Company for a Form Field: Why One Main Agent Beats a Full Roster
Multi-agent orchestration adds token usage, handoff overhead, and latency to tasks a single agent can finish in a continuous loop. Teams that copy corporate role structures onto agents pay for process they don't need, while those that follow Anthropic's fold-in guidance cut cost and time without losing quality—and the article gives concrete criteria for when splitting actually pays off.
Multi-agent coding workflows that mirror a company org chart—product, architecture, development, testing, plus a coordinator—cost more in tokens, handoff overhead, and latency than they save on typical tasks like adding a field to a form. Anthropic's guidance on folding subagents into the main loop reports that some enterprise customers halved elapsed time after switching to one main agent that loads Skills and tools on demand, with quality improving too.
Handoff costs don't disappear just because agents don't complain. When implementation and testing are split across agents, the tester has to reconstruct why a compatibility handling exists—was it a product requirement or historical debt—or re-read the code. Writing fuller handoff documents is itself work, and judging which exploration results are worth preserving is part of the job. Anthropic's multi-agent architecture guidance similarly suggests dividing along isolable contexts, keeping implementation and testing of the same feature together.
The alternative is to encode working methods into Skills: impact analysis before code changes, check verification on delivery, request wrapper conventions. A main agent can then follow those steps without spawning an analyst. Splitting into separate agents is worth it only for independent exploration (multiple root causes, unrelated modules) and independent review where a second agent checks the final artifact against requirements. Built-in subagents in Codex and Claude Code handle those cases without a full orchestration system, but team modes burn more tokens and fit sequential, same-file, or dependency-heavy work poorly. The test is whether the main agent can do useful work while the subagent runs.
Agent architectures are mirroring corporate hierarchies without inheriting the reasons those hierarchies exist: humans differ in experience and have limited time, but agents built on the same model don't gain different expertise from a title.
Handoff costs that are grudgingly accepted in human teams are silently assumed to be free between agents, yet judging which exploration results are worth documenting is itself real work that the handoff document can't fully offload.
Confusing "needs a different method" with "needs a different agent" is the trap; a Skill that stipulates impact analysis before editing code removes the need for an analyst role entirely.
Whether a split pays off reduces to whether the parent agent can do useful work while the child runs; if it just waits and then rereads code, the extra division didn't save time.