Apodex 1.1: Scaling Agentic Intelligence for Complex Work

The Core Claim

Language models can talk intelligently, but real work requires them to do things reliably over time—interact with files, search databases, run code, handle failures, and prove they've made progress. Apodex 1.1 demonstrates that a moderately-sized model (35 billion parameters) can execute complex professional tasks at a level competitive with much larger systems by training it to coordinate across diverse tools and manage long-running workflows with state recovery.

How It Works

The authors built their agent along two tracks:

Environment Scaling means expanding what tools the agent can touch and how reliably it can verify results. Instead of a generic "execute code" primitive, the system includes diverse executables—file operations, web search, calculators, code sandboxes—each with clear input/output contracts. The key insight is verifiability: the agent learns not just to call tools but to check that its actions actually produced the intended results.

Agentic Coordination Scaling trains the agent to manage complexity over long horizons. Rather than running a single chain of thought, it learns to break tasks into subtasks, run some in parallel (asynchronously), collect results from multiple workers, and replan when things fail. This mirrors how humans tackle complex projects: divide, delegate, wait, integrate, adapt.

Underlying both is an AgentOS—a kernel that maintains task state, tracks where each result came from (provenance), and lets the system recover gracefully when things go wrong. The shared execution harness logs everything, turning the system's journey through a task into training data. The authors then use these real trajectories to teach the agent which decompositions work, which tool calls succeed, and how to coordinate effectively.

Why It Matters

The results are striking because they're useful. Tested across six domains—complex finance analysis, scientific research, math, coding, general search—Apodex 1.1 hits the "leading performance band" despite using a 35-billion-parameter model. Competitors often rely on 70B+ or proprietary models. The 35B Mini variant runs locally, removing cloud dependency for deployments.

More fundamentally, the paper grounds "agentic intelligence" in something concrete: not better chat, but verifiable progress on real work. A finance analysis must produce defensible numbers; a scientific task must cite sources; code must run. The system's ability to maintain state across time, recover from failures, and prove what it did aligns AI capability with what humans actually need from tools. This matters because previous agentic systems often falter on long tasks, hallucinate tool calls, or lose track of context.

The Heavy-Duty Solver framing is intentional: the authors are targeting sustained, ambitious projects—the kind that take hours or days, not seconds—rather than one-shot completions.