AI Insights · Oct 22, 2026 · Verified

When to split work across agents

Separate agents work well for research, review and independent checking. Several agents writing the same code in parallel is still unproven.

Claim

Spawn subagents to isolate context for reads (research, exploration, review) and for adversarial verification with no shared prior context. Keep writes single-threaded or partition them with an oracle. Generic peer "crews" and shared-context swarms do not work.

Evidence

  • Anthropic research system (2025-06-13): lead plans and saves the plan to memory, parallel subagents with own context return 1k to 2k-token condensed results; 90.2% better than single-agent at about 15x chat tokens; token usage explains about 80% of BrowseComp variance; delegation needs explicit objective, format, tool guidance and boundaries; coding has fewer parallelisable tasks.
  • Cognition (2025-06-12): actions carry implicit decisions, so share full traces and keep writers single-threaded. Follow-up (2026-04-22): parallel writer swarms still lack adoption; what works is generator-verifier review with no shared prior context (Devin Review), "smart friend" escalation to a frontier model, and map-reduce with one synthesising writer.
  • Anthropic C compiler (2026-02-05): 16 parallel instances, about 2,000 sessions, just under $20,000, no orchestrator; tasks claimed via lock files in git; parallelism failed on one shared task until GCC was used as an oracle to partition files; "test suites must be nearly perfect".
  • LangChain docs: a single agent with the right dynamic tools often matches multi-agent; the strong reasons to go multi-agent are context management, distributed development and parallelisation.

Open questions

  • Whether verifier agents should see the plan or only the output.

Sources

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.

Related

Book a demo