Source-linked AI summary
From Skills to Talent: Organising Heterogeneous Agents as a Real-World Company
Zhengxu Yu, Yu Fu, Zhiyuan He, Yuxuan Huang, Lee Ka Yiu, Meng Fang, Weilin Luo, Jun Wang
TL;DR
Multi-agent systems lack an organisational layer for coordinating, managing, and evolving heterogeneous agents beyond individual skills and fixed interaction structures. OMC addresses this gap with portable Talents, typed Containers and interfaces, on-demand recruitment, E2R tree search, and self-evolution mechanisms. It achieves an 84.67% success rate on PRDBench, while case studies provide initial evidence of cross-domain generality.
Problem
Existing multi-agent systems remain limited by hardcoded or unconstrained team coordination, incompatible runtimes, descriptive roles, and session-bound self-improvement.
Method
OMC combines portable Talent identities, typed Talent-Container interfaces, a community-driven Talent Market, E2R tree search, and agent-organisation self-evolution.
Results
84.67% success rate on PRDBench surpassed all baselines by at least 15 percentage points in a single-attempt zero-shot setting.
Takeaways & Limitations
OMC demonstrates an organisation-level approach for structuring, coordinating, and evolving heterogeneous agent workforces, with initial cross-domain evidence from case studies.
Takeaways & Limitations
Quantitative evaluation is confined to 50 PRDBench software-development tasks, and the self-evolution mechanisms have not yet been quantitatively ablated.
Abstract
from arXiv · showhide
Individual agent capabilities have advanced rapidly through modular skills and tool integrations, yet multi-agent systems remain constrained by fixed team structures, tightly coupled coordination logic, and session-bound learning. We argue that this reflects a deeper absence: a principled organisational layer that governs how a workforce of agents is assembled, governed, and improved over time, decoupled from what individual agents know. To fill this gap, we introduce \emph{OneManCompany (OMC)}, a framework that elevates multi-agent systems to the organisational level. OMC encapsulates skills, tools, and runtime configurations into portable agent identities called \emph{Talents}, orchestrated through typed organisational interfaces that abstract over heterogeneous backends. A community-driven \emph{Talent Market} enables on-demand recruitment, allowing the organisation to close capability gaps and reconfigure itself dynamically during execution. Organisational decision-making is operationalised through an \emph{Explore-Execute-Review} ($\text{E}^2$R) tree search, which unifies planning, execution, and evaluation in a single hierarchical loop: tasks are decomposed top-down into accountable units and execution outcomes are aggregated bottom-up to drive systematic review and refinement. This loop provides formal guarantees on termination and deadlock freedom while mirroring the feedback mechanisms of human enterprises. Together, these contributions transform multi-agent systems from static, pre-configured pipelines into self-organising and self-improving AI organisations capable of adapting to open-ended tasks across diverse domains. Empirical evaluation on PRDBench shows that OMC achieves an $84.67\%$ success rate, surpassing the state of the art by $15.48$ percentage points, with cross-domain case studies further demonstrating its generality.
1 Introduction
OMC frames AI organisation as a missing layer for structuring, coordinating, managing, and evolving heterogeneous agent workforces beyond individual skills and conventional multi-agent interaction. It introduces portable Talents, typed execution abstractions, on-demand recruitment, E2R-based coordination, and feedback-driven evolution, achieving 84.67% success on PRDBench.
- Research gap: Existing multi-agent systems hardcode teams or permit unconstrained negotiation, while incompatible runtimes, descriptive roles, and session-bound improvement limit workforce adaptability.These limitations motivate an organisational layer that governs heterogeneous agents beyond individual skills and interaction protocols.
- Research gap: An AI organisation is a self-governing system of heterogeneous agents with structured coordination, managed lifecycles, and experience-driven evolution.Its organisation-level abstraction governs workforce assembly, work decomposition, execution, management, and improvement over time.
- OMC framework: OMC separates portable agent identities from execution environments through Talents and Containers, using typed organisational interfaces across heterogeneous backends.Talents package roles, prompts, skills, tools, and working principles; Containers host them on runtimes such as LangGraph, Claude Code, or script processes.
- OMC framework: The Talent Market recruits verified agents on demand, while lifecycle management supports provisioning, evaluation, and potential offboarding.This elevates reuse from individual skills to complete agent identities with managed lifecycles.
- OMC framework: Explore-Execute-Review tree search decomposes projects, executes work through agents, and reviews outcomes using DAG-based task execution with termination and deadlock-freedom guarantees.The mechanism uses AND-tree semantics and a finite state machine to support structured coordination and crash recovery.
- Evaluation: 84.67% success rate on PRDBench surpassed all baselines by at least 15 percentage points in a single-attempt zero-shot setting.The evaluation uses a project-level software development benchmark.
2 Methodology
OMC treats multi-agent organisation design as a first-class layer for recruiting, coordinating, executing, reviewing, and evolving heterogeneous agents. Its Talent–Container abstraction, Talent Market, E2R search, DAG execution, bounded controls, and lifecycle processes support dynamic organisational operation.
- Framework overview: OMC constructs, manages, and evolves multi-agent organisations through an organisational layer, Talent Market, E2R search, DAG scheduling, reflection, and performance review.The framework is designed as an organisation rather than a fixed agent with a predefined workflow.
- Employee composition: Talent packages define portable agent identities, while Containers provide heterogeneous runtimes and typed organisational interfaces for executing them.The same Talent can run across supported backends without modification, while a Container hosts the runtime and organisational contracts.
- Talent Market: The Talent Market recruits community-verified, benchmark-validated agent packages and can rank candidates by skill match and community ratings when capabilities are missing.Each marketplace Talent includes executable prompts, tools, skills, domain resources, and documented benchmark results.
- E2R execution: E2R explores decomposition strategies, executes assigned work, and reviews outcomes, while DAG semantics propagate accepted subtask results bottom-up toward project completion.The system can reject results and explore a new decomposition under the same parent using updated context.
- External oversight: Human stakeholders can override policies, inject requirements, and trigger or stop iterations, but convergence depends on the quality of human judgment rather than formal guarantees.This design prunes unpromising branches and supplies external information unavailable to the system.
- Execution guarantees: Bounded review rounds, task timeouts, and cost budgets guarantee termination of each search episode in bounded time and cost when executors respect the timeout contract.The review limit defaults to k_rev = 3 and the task timeout defaults to T_max = 3600s.
- Self-evolution: Lifecycle management links recruitment and organisational evolution by replacing underperforming agents and allowing high-performing agents to accumulate experience.The paper frames structured HR protocols such as performance reviews, PIPs, and formal offboarding as part of agent management.
3 Experiments
OMC is evaluated on PRDBench and across four case studies, showing strong software-development performance and cross-domain operation through dynamic recruitment and task execution.
- PRDBench contains 50 project-level software-development tasks spanning over 20 domains, with structured requirements and executable evaluation scripts.
- 84.67% success rate (+15.48%) is the highest across systems on PRDBench, surpassing all baselines.The evaluation uses the official DEV mode, with one-shot inputs and automated final evaluation.
- OMC dynamically adjusts decomposition, enforces supervisor approval, and matches heterogeneous backends to subtasks during execution.These mechanisms are identified as contributing to the reported result.
- The game-development case uses evaluator feedback to create a new Art Designer skill and re-execute the asset pipeline after rejection.
- The research-survey case completed in under one hour at $16.26, producing 17 structured documents, a six-theme mind map, and three research proposals.The outputs were produced in a single zero-shot iteration and manually verified for cited-paper reality, taxonomy structure, and novelty of one proposal.
- Across four case studies, OMC recruits specialists, decomposes projects, executes across heterogeneous backends, and delivers results without domain-specific framework changes.The cases cover content generation, game development, audiobook production, and academic research.
4 Related Work
Related work is compared across heterogeneous-agent abstraction, dynamic coordination, and persistent self-improvement, with OMC addressing identified architectural gaps.
- The literature is organised around unified management of heterogeneous agents, dynamic task execution, and persistent self-improvement.
- Table 4 compares design paradigm, execution model, agent contract, state management, multi-family support, sourcing, and evolution capabilities.
- Agent Heterogeneity and Runtime Abstraction: Most frameworks couple agent identity to one execution model, while OMC uses six typed organisational interfaces to decouple identity from runtime.
- Dynamic Task Coordination: Existing systems generally either hardcode workflow graphs or adapt decomposition without formal completion guarantees; OMC combines E2R, DAG execution, and FSM lifecycle control.
- Self-Evolution and Organisational Evolution: No compared system implements both individual and organisational self-evolution, whereas OMC uses reflection, retrospectives, and HR performance reviews without model retraining.
5 Discussion
The discussion establishes evaluation and deployment boundaries: quantitative evidence is limited, self-evolution lacks ablation, and coordination adds cost that may not suit simple queries.
- Quantitative evaluation is confined to PRDBench’s 50 software-development tasks, while systematic non-coding benchmarks remain future work.Cross-domain case studies provide initial evidence of applicability but do not replace systematic non-coding evaluation.
- OMC’s one-on-ones, retrospectives, and performance reviews have been deployed but not quantitatively ablated.Isolating each mechanism’s contribution requires longitudinal studies across many projects.
- Approximately $6.91 per PRDBench task reflects multi-agent coordination overhead, making OMC less appropriate for simple single-turn queries.The framework therefore includes adaptive dispatch to route simple tasks to one agent and reserve coordination for tasks above a complexity threshold.
- The manage–plan–hire–learn pattern is presented as domain-general, with four case studies supplying initial evidence beyond software development.
6 Conclusion
OMC frames AI organisation design as a missing layer for structuring, coordinating, and evolving heterogeneous agent workforces. Its Talent–Container architecture, organisational interfaces, E2R search, and self-evolution pipeline support this organisational model.
- OMC combines Talent–Container workforce management, E2R execution coordination, and self-evolution through reflection, retrospectives, and HR processes.
- 84.67% success on PRDBench surpassed all baselines across multi-domain, long-horizon projects.
- Organisational interfaces map agent-workforce functions to classical OS kernel subsystems.
- The interfaces cover process creation and scheduling, task dispatch and mutual exclusion, runtime context assembly, storage, event publication, communication, lifecycle, security, and audit.
- Containers abstract backend-specific APIs so the platform can govern heterogeneous agent runtimes through common organisational contracts.
C Executor Backends and Task Dispatch
OMC separates task dispatch, execution, permissions, and coordination from individual executor backends. This supports stateless heterogeneous runtimes, persistent recovery, and uniform organisational tool access.
- On-demand dispatch starts assigned tasks immediately, avoids idle resource use, and recovers interruptions through persisted state.
- Three reference backends support LangChain reactive agents, Claude Code sessions, and script-based executors through a decoupled execution interface.
- Failures are retried with injected error context and escalated for supervisory review after a configurable number of attempts.
- Role-based permissions restrict each agent’s available tools before reasoning begins, while MCP exposes organisational tools across executor backends.
- Inter-agent coordination is expressed through tools for delegation, subtask decisions, discussion, and CEO escalation.
D Talent Market: A Three-Type Agent Supply Chain
The Talent Market supplies portable agent packages through three parallel sourcing channels, all entering OMC through the same Container-governed runtime. Feedback from deployment can enrich the community supply pool.
- The Talent Market determines what cognitive capability is available for hire, complementing Containers’ role in defining execution.
- All sourcing types produce standard Talent packages loaded into Containers, and hired agents begin at the same employee level.
- Type 1 packages validated agents from established open-source repositories while removing framework-specific runtime dependencies.
- Type 2 assembles agents from community-curated specialist personas that specify identities, principles, deliverables, and success metrics.
- Type 3 constructs personas and skill sets on demand from modular cloud skills retrieved through SkillsMP.
- The CEO makes the final selection from a ranked top-k shortlist, approximately 80% Type 1 and Type 2 candidates and 20% Type 3.
- Deployed-agent feedback can publish refined skills and effective synthesized personas back to skill and prompt repositories.
E Case 1: GitHub AI Agent Weekly Trend Report
The case study reports an autonomously produced weekly trend report workflow involving research, technical writing, and CEO delivery without human editing.
- The Researcher agent collected research data that the Writer agent used to produce the report.
- The Writer agent autonomously produced the report using Claude Sonnet 4.
- The completed report was emailed to the CEO without human editing.
GitHub AI Agent Trending Repositories Report
The report surveys 15 AI-agent repositories that gained traction from March 18–25, 2026, highlighting cross-platform integration, agent collaboration, specialised applications, and ecosystem growth.
- 15 notable repositories spanning enterprise and specialised applications accumulated over 4,000 stars during March 18–25, 2026.
- Key Market Trends: Cross-platform integrations, particularly for Chinese platforms, emerged as the most significant trend, including WeChat and AstronClaw/Loomy projects.
- Key Market Trends: Agent-to-agent collaboration is represented by ClawLink’s communication network and HyperAgents’ self-referential self-improvement.
- Key Market Trends: Specialised applications include App Store compliance scanning, autonomous CTF solving, and business-development automation.
- Conclusion: The report attributes 4,000+ stars across 15 repositories to strong developer interest and continued rapid growth in the AI-agent ecosystem.
F Case 2: Game Development
The game-development case study assembles a Game Developer and Art Designer to build a street-fighting web game, with external human evaluation demonstrating human-in-the-loop iteration.
- Figure 10 documents team assembly, recruited Talent profiles, collaboration, generated project documents, and cost breakdown.
- A Game Developer and Art Designer collaborate to develop a street-fighting web game within OMC’s execution framework.
- The workflow includes external evaluation by a human tester, demonstrating human-in-the-loop iteration.
G Case 3: Audio Book Development
The audiobook case study coordinates a Novel Writer and AV Producer across screenplay, illustration, voice-over, and video-assembly stages using different LLM backends.
- The case study figure reports approximately 1.56M tokens costing $1.57, or 15.7% of the $10 budget.
- A Novel Writer and AV Producer create an illustrated audiobook retelling of Peaky Blinders with animal characters.
- The sequential workflow covers screenplay adaptation, scene illustration, voice-over synthesis, and final video assembly.
- The two agents use Claude Sonnet 4 and Gemini 3.1 Pro while coordinating across distinct media pipelines.
H Case 4: Research Survey - Team and Generated Ideas
The research-survey case study assembles three specialists to produce survey artefacts and autonomously generate research ideas grounded in identified world-model failure modes.
- Team and Generated Ideas: Three recruited specialists—two Research Scientists and one AI Engineer—produce 18 deliverable documents autonomously.
- Team and Generated Ideas: The survey team’s outputs include literature reviews, failure-mode taxonomies, and research proposals.
- Generated Ideas: The first idea addresses compounding prediction error with a fast short-horizon model, slow macro-state model, and uncertainty-gated re-grounding.
- Generated Ideas: The second idea injects differentiable physics priors into latent transitions to address violations of conservation laws and contact constraints.
- Generated Ideas: The third idea combines meta-learning across randomized simulation domains with conformal prediction to address sim-to-real shift and hallucination.
- Survey Artefacts: The mind map covers six themes with approximately 70 nodes referencing 35+ papers from 2021–2026.