The Big Picture
A prompt-driven, no-code multi-agent system executed over 1,000 research iterations across more than ten fields, automating experiment code and boosting iteration throughput by an order of magnitude while keeping humans as the final decision-makers.
ON THIS PAGE
The Evidence
A reusable, prompt-first architecture lets many research threads run in parallel while keeping prompts short and model-specific details out of the control logic. The system produced thousands of scientist–coder–auditor interactions with zero hand-written experimental code and the same orchestrator transferring across fields without modification. A producer–critic loop around each key artifact (idea, proposal, experiment, paper) catches many errors, and a failure-mode taxonomy shows where automation stops and human judgment is still required. This aligns with a Planning Pattern for organizing concurrent work and orchestrating handoffs.
Data Highlights
1Executed over 1,000 scientist–coder–auditor iterations across more than 10 scientific domains.
2Zero human-written experimental code across deployments (a fully zero-code experimental pipeline).
3Throughput claim: runs hundreds of experimental iterations in the time a human lab runs ten (≈10x–100x higher iteration rate).
What This Means
Engineers building multi-agent research orchestration and platform teams looking to automate repetitive experiment loops will gain a blueprint for reusable prompts and artifact-based handoffs. Lab leaders and technical managers can use the failure-mode taxonomy to decide which parts of their workflow can be automated safely and where human oversight must remain. See examples in the Role-Based Agent Pattern.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1: System overview of Agon . The workflow starts from either topic-radar or human topic selection, then proceeds through idea, proposal, experiment, and paper factories. Each factory advances a research artifact through role-specific agent loops, while deep-literature research supplies reusable context to multiple stages.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Results reflect deployments within a single research group's infrastructure and depend on current generation foundation models, so performance will shift as models evolve. Several failure modes remain invisible to the agents and require human experts for detection and judgment. Ethical oversight, reproducibility checks, and domain-specific validation remain necessary before using outputs as publishable science or production decisions. For common failure modes, see Inter-Agent Miscommunication.
Methodology & More
Agon is built around six practical principles: keep prompts small and reusable (prompt economy), avoid baking model-specific details into control prompts (future-facing), use minimal prompts and zero code, design to span many disciplines, and run massive parallel threads. The system organizes work into 'factories'—loops that produce and refine artifacts (ideas, proposals, experiments, papers)—paired with independent critics that try to break the artifact before it advances. A prompt-native dispatcher manages many concurrent projects and mediates handoffs by persisting artifacts in a versioned repository, which makes the workflow auditable and transferable across topics. Deployment evidence shows the same orchestrator ran more than 1,000 scientist–coder–auditor iterations across 10+ domains without any human writing experimental code, demonstrating cross-disciplinary transfer and large throughput gains (reported as hundreds of iterations versus a human lab’s ten). The team also created a failure-mode taxonomy along dimensions of cost, fixability, visibility, and locus of capability; the key limiting factor for automation is visibility—failures invisible to the loop require human insight. Agon is open-source and intended as a practical blueprint for democratizing parts of research while explicitly marking the boundary where human steering remains essential. For practical integration patterns, consider the Tool Use Pattern.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
ArXiv preprint with no affiliations or known author reputations listed.