Most LLM research focuses on single-turn reasoning or short tool-assisted tasks. Real-world complex work instead needs sustained, verifiable progress across files, search, code, and multiple agents — including state maintenance, asynchronous delegation, failure recovery, and end-to-end verification. This paper frames and advances that "working capability" by scaling both environments and agentic coordination so models can complete multi-step professional tasks reliably over time.
Key Findings
- Environment scaling: Expands the diversity and verifiability of executable environments (files, search, code) so agent behaviors can be executed and audited, not just speculated in-context. This increases reproducibility and enables training from real execution trajectories.
- Agentic coordination scaling: Trains agents to decompose long-horizon tasks, delegate parallel work, integrate asynchronous results, replan, and preserve provenance — shifting behavior from single-agent loops to orchestrated multi-agent workflows.
- Shared execution harness (AgentOS): A runtime that maintains task state, checkpoints, and provenance across tools and agents, letting the system recover from failures and perform global verification and repair before delivery.
- Empirical outcome: The system attains top-tier performance on diverse professional, scientific, mathematical, coding, and search workloads while using smaller base models; a 35B-parameter Apodex 1.1 Mini preserves strong working capability suitable for local deployment.
Who it's for and trade-offs
Great fit if you need AI to complete multi-step, auditable work that interacts with files, executables, and external search — e.g., research assistants, engineering tasks, finance workflows, or long-running coding projects. The approach favors systems where verifiability, provenance, and recoverability matter as much as raw single-step accuracy. Look elsewhere if your needs are short-horizon, latency-sensitive inference with minimal external state or if you require turnkey cloud-hosted closed-source agents; the focus here is on robust long-horizon execution and training infrastructure rather than lightweight conversational APIs.
Where it fits
Positions itself between single-agent LLM tool-using systems and fully hand-engineered workflow engines: it combines environment-level scaling (to make executions inspectable and trainable) with coordination-level scaling (to orchestrate distributed agent work), aiming to provide a practical "heavy-duty solver" for ambitious, verifiable tasks.