Apodex 1.1 Ships: Agent System Measured by End-to-End Deliverables

Apodex rolled out version 1.1 and its ecosystem components in rapid succession on September 8–9: the open-source agent framework FrontierAgent is now on GitHub with around 2.4k stars, mini weights (including FP8, NVFP4, and GPTQ-Int4 quantized versions) are available on Hugging Face for local deployment, the full model powers an online web workspace, and the accompanying API is free for new users for two weeks. The entire release rests on the claim that "even a correct answer can be a failed task": the meaningful unit for measuring AI capability is completing a real piece of work end to end and producing verifiable deliverables such as files and code.

Confirmed

• The FrontierAgent framework runs locally with a single command, features a native command-line TUI, and supports both single-agent and Agent Team modes (m1)
• The Apodex 1.1 paper (arXiv:2608.23283) and its Agent Team evaluation results were added to Papers with Code by Niels Rogge (m2)
• Official figures place Agent Team mode among the current frontier tier of agent systems on benchmarks including APEX-Agents, GDPval, FrontierScience, and Humanity's Last Exam; FrontierScience-Research pass rate of 63.3% (m2, m5)
• Demos handle real professional files directly: reading raw clinical data to complete cleaning, method selection, and code execution while producing Kaplan-Meier curves; other demos cover powder XRD indexing and batch protein docking (m6)

Core Mechanisms

• Statement Review: generation and review are separated; important conclusions undergo independent review before delivery; when evidence is insufficient, citations don't match, or numbers conflict, the system flags the issue, shows corrections, and retains the review record (m7)
• Mid-run intervention: the asynchronous Agent Team supports injecting new files or changing requirements during execution, preserving still-valid results and re-planning only affected parts; the official demo covers risk assessment for a global EV battery supplier (m8)
• Deep Discover engine: the model decides for itself how to decompose tasks, how many sub-agents to spawn, and when to aggregate; sub-agents run in parallel and stream results back to a shared task state (m9)

Why it matters

@aakashgupta noted across several reviews that 1.1 shifts the evaluation unit from "answering questions correctly" to "delivering verifiable results," and its built-in independent verification mechanism (Statement Review) addresses the credibility concerns of using agent outputs in real professional settings; the locally deployable mini 35B plus a two-week free API also lower the barrier to actually trying it out.

2026-09-08 ~ 2026-09-09 · 10 related posts

• Apodex 1.1 Agent Team hits 63.3% pass rate on FrontierScience-Research, paper lands on Papers with Code — NielsRogge · 2026-09-08
• [source] Apodex 1.1 Argues a Right Answer Can Still Be a Failed Task — aakashgupta · 2026-09-09
• Apodex 1.1 argues a right answer can still be a failed task — aakashgupta · 2026-09-09
• Inside Apodex: an asynchronous Agent Team running on a live task board — aakashgupta · 2026-09-09
• Apodex agents absorb mid-task requirement changes with a partial replan — aakashgupta · 2026-09-09
• Apodex 1.1 review: Statement Review independently verifies agent conclusions — aakashgupta · 2026-09-09
• Apodex 1.1 Demos Hit Real Clinical Data, and Statement Review May Be the Sleeper Feature — aakashgupta · 2026-09-09
• Apodex Claims 1.1 With Agent Team Sits in the Frontier Tier; 35B Mini Runs Locally — aakashgupta · 2026-09-09
• [source] Apodex 1.1 Mini Weights Hit Hugging Face; Full Model API Free for Two Weeks — aakashgupta · 2026-09-09
• [source] Apodex Open-Sources FrontierAgent: One-Command Local Agent Framework With Agent Team Mode — aakashgupta · 2026-09-09