Five Gemini 3.5 Flash agents working together in real time solved 87 % of a 30-question Project Euler set, beating five agents running solo (72 %) and both majority-vote baselines (80 % solo, 90 % collaborative). The collaborative run also shaved four minutes off the average execution time, dropping from 14 minutes per problem to 10 minutes, with most solutions arriving in under five minutes.

Why the experiment matters

AI coding assistants already draft snippets, debug code, and generate whole programs. Researchers usually test a single model on a prompt and grade its answer. This test asks a different question: can a group of agents that share findings as they work outperform a “wisdom of the crowd” approach that merely tallies independent answers?

Project Euler problems serve as a standard benchmark for algorithmic thinking. They mix mathematical insight and efficient coding, making them a good proxy for real-world programming tasks where correctness and speed matter.

How the test was set up

  • Agents: Five instances of Gemini 3.5 Flash accessed through the Antigravity platform.
  • Scenarios:
    1. Each agent tackled every problem alone, reporting its own answer.
    2. The same five agents collaborated in real time, broadcasting intermediate results and confirming each other’s steps.
    3. For both setups, a simple majority-vote aggregator combined the five independent answers.
  • Metrics: Accuracy (percentage of correct answers) and runtime (average time per problem, plus the distribution of completion times).

What the numbers reveal

  • Solo accuracy: 72 %
  • Collaborative accuracy: 87 %
  • Solo majority-vote accuracy: 80 %
  • Collaborative majority-vote accuracy: 90 %

Runtime dropped from 14 minutes on average for solo agents to 10 minutes for the collaborative group. Most collaborative runs finished under five minutes, whereas solo attempts frequently hit the 30-minute timeout limit.

Key observations that explain the gap

  • Impossible for one, possible for many – On a single problem all solo agents failed, yet the collaborative team exchanged partial results and collectively arrived at the correct answer.
  • Error catching through cross-checking – Individual agents sometimes fell into logical traps; the group caught these missteps by comparing notes in real time.
  • Rapid convergence – When any agent discovered a crucial intermediate value, it broadcast the finding, allowing the others to skip redundant work and converge on the final solution faster.

The hidden costs and limits

The experiment used only five agents; it remains unclear whether the same efficiency scales to dozens or hundreds. Moreover, the majority-vote aggregator still performed well (90 % with collaboration).

What to watch next

  • Scaling experiments – Will a hundred agents still improve accuracy, or will coordination overhead dominate?
  • Role specialization – Assigning specific tasks (e.g., one agent focuses on number theory, another on optimization) could amplify benefits over free-form collaboration.
  • Broader benchmarks – Applying the same framework to code-generation challenges, debugging suites, or real-world software builds will test whether the gains hold beyond mathematical puzzles.

Takeaway

Real-time collaboration among AI coding agents can raise both success rates and speed on algorithmic tasks, outperforming solo attempts and even a simple crowd vote. The approach shows promise, but its scalability and cost-effectiveness remain open questions that future research will need to answer.

Source: https://dev.to/ykdap/for-coding-agents-real-time-collaboration-beats-the-wisdom-of-the-crowd-1kj0