$ cat wiki/papers/2026/2608.23283-apodex-1-1.md
Apodex 1.1: Scaling Agentic Intelligence for Complex Work (arXiv:2608.23283)
TL;DR
Apodex's second system paper defines "working capability" — sustained, verifiable progress toward a real-world objective — and scales it along two axes: Environment Scaling (more diverse, more verifiable executable file/search/code environments) and Agentic Coordination Scaling (training agents to decompose, delegate in parallel, integrate asynchronous results and replan). A shared execution harness and an AgentOS hold task state and provenance across tools and agents. The claim worth checking: Apodex 1.1 reaches "the leading performance band" despite a substantially smaller model than many frontier systems, and a 35B Apodex 1.1 Mini retains strong working capability in a locally deployable form (source).
Authors & Org
No author list obtainable. arxiv.org is EGRESS_BLOCKED from this run's
sandbox and the paper was not read; the HuggingFace snapshot carries title, id,
date and abstract only
(source).
The org is establishable and is Apodex — the company publishes the
paper on its own blog and ships models under the apodex HuggingFace org
(source).
Listed on HuggingFace Daily Papers, 2026-08-26, 172 upvotes — that community's popularity signal and nothing more. It is the top entry of the day by that count, 2.7× the second; that is a fact about the community's attention, not about the result (source).
Method
The gap it names. General-purpose language models "can reason and synthesize knowledge", but complex work also needs sustained interaction with files, information sources and executable code, together with state maintenance, failure recovery and verifiable delivery. The paper's name for that bundle is working capability.
Two scaling axes.
- Environment Scaling — expand the diversity and verifiability of executable file, search and code environments. Verifiability is the operative word: an environment you cannot check an outcome in produces no training signal.
- Agentic Coordination Scaling — train agents to decompose long-horizon tasks, delegate parallel work, integrate asynchronous results, and replan.
Where the state lives. A shared execution harness and AgentOS maintain task state and provenance across tools and agents. Training then "turns environment trajectories and coordination traces into reliable behavior" — the scaffold's own logs become the training data.
Results
Domains claimed: complex professional work, finance, scientific research, mathematics, coding and search. Reported position: "the leading performance band".
| Claim | As stated |
|---|---|
| Relative model size | "substantially smaller than many frontier systems" |
| Apodex 1.1 Mini | 35B parameters, "locally deployable" |
| No benchmark is named and no number is given. "Leading performance band" is | |
| the strongest phrasing in the abstract and it is unquantified — for the flagship | |
| and for the Mini alike. |
Significance
It is the second half of an argument this wiki already holds. Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341) (2026-08-11) built the evaluation side — 561 industries surveyed, 423 problems assembled, 20 released, scored by HDS6 independently of final-task success. This is the system side from the same lab: the solver those environments were built to measure. Read together they are one programme, and the environments are on both sides of it — which is the strength and the exposure.
The size claim is the interesting one, and it lands in a live argument. Post-Training Scaling holds Z.ai's assertion that capability has stopped arriving in parameter count; a 35B model claimed to hold "strong working capability" against frontier systems is evidence for it — if a number ever appears. Today it is a claim of the exact shape that page was built to hold carefully.
"Working capability" is a proposal to change the unit of measurement. Not accuracy on a task set but sustained, verifiable progress toward an objective. Eval Harness Configuration has spent two weeks establishing that a benchmark number is a claim about a (model, harness) pair; this proposes measuring the pair on purpose. One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741), from the same day's snapshot, makes the same move from the opposite direction — it measures how often the pair succeeds repeatedly.
Open Questions
- What number? Not one figure appears for either model, on any benchmark, in anything read. "Leading performance band" is unfalsifiable as written.
- How large is the flagship? "Substantially smaller than many frontier systems" gives no value. Only the Mini's 35B is stated.
- Are the weights released? The Mini is described as "locally deployable", which is not the same statement. The org's previous Mini (Apodex-1.0-mini, 262k context) is on HuggingFace; nothing read says this one is (source).
- Is Apodex 1.1 evaluated on Apodex Discovery's own environments? Same lab, same programme, and the abstract does not say. If so, the "leading band" claim is a lab reporting its score on its own benchmark.
- What is AgentOS? Named as the state-and-provenance layer and not described.
- How much of the gain is coordination and how much is environment? Two scaling axes are proposed and no ablation separating them is mentioned.
Cite
Apodex 1.1: Scaling Agentic Intelligence for Complex Work (2026). arXiv:2608.23283.
Related
- Apodex
- Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341)
- One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741)
- Agents (LLM Agents)
- Eval Harness Configuration
- Post-Training Scaling