Cohere researcher Nouha Dziri studies how to allocate work across coding-agent teams in this 31-minute AI Engineer research talk.

Find the bottleneck before adding workers

  • A lead assigns narrow tasks to workers with fresh context, then combines their findings in a shared workspace.
  • On ProgramBench, agents reconstruct software from a black-box executable and documentation; discovering its behavior is the hard part.
  • Planning workers probe different interfaces and edge cases while the lead handles implementation and verification.
  • On Commit0, repository skeletons and tests already supply much of the plan, so implementation workers help more.

Coordination changes the result

  • With minimal instructions, teams improved benchmark scores but cost roughly four to seven times as much as a solo agent.
  • An asynchronous lead sometimes acted before reports arrived or submitted while workers were still running.
  • Explicit instructions to collect reports closed much of that coordination gap.
  • Planning-only delegation matched the broader plan–implement–verify team’s reward at about half its cost on ProgramBench.

Read the 70% saving against its actual baseline

  • A strong lead with cheaper planning workers approached an all-strong team’s score at lower inference cost.
  • The research paper specifies one strong lead plus 16 cheaper planners, within 1.4 score points of the comparison team at 70% lower cost.
  • That comparison is against an expensive homogeneous team delegating across stages, not a blanket discount versus a solo agent.
  • Extra implementation and verification workers plateaued in this setting; weaker leads also limited what their workers could achieve.

More messages can add overhead

  • In the tested teams, 84% of messages went from workers to the lead, with fewer than one peer-to-peer message per task.
  • Adding messaging left stronger teams’ scores roughly unchanged and reduced weaker teams’ scores.
  • Dziri treats this as task- and infrastructure-dependent, leaving larger teams, shared awareness, and safer coordination as open research questions.

Caveat: these are the researchers’ benchmark results, with extensive prompt tuning and a specific orchestration setup; general coding workloads may bottleneck elsewhere.

Research paper: Scaling Inference-Time Compute with Teams of Coding Agents supplies the precise comparison behind the headline.

“The lead sets the ceiling and the workers set the price.” — Nouha Dziri, 21:45