Apodex 1.1 Shifts from Research Q&A to Real Task Execution
Apodex released its 1.1 model family on September 16 (TestingCatalog reported it went live on August 24), repositioning from "deep research" Q&A to end-to-end execution of long-horizon real-world tasks, prompting follow-up deep dives from multiple authors.
Confirmed
- According to @testingcatalog, Apodex 1.1 can directly open uploaded files (spreadsheets, PDFs, datasets), choose analysis methods, and write and run analysis code, ultimately returning traceable tables, charts, and scripts—every conclusion in the deliverables can be traced back to its original source.
- @aakashgupta relayed its core claim: a correct answer can still count as a failed task. The meaningful unit of AI capability is completing real work end-to-end—including six things like understanding goals, manipulating files, writing code, and producing verifiable results—not merely generating a plausible-looking report.
- @aakashgupta focused on testing mid-task intervention: when new files are injected or requirements change during a running task, the system preserves work that remains valid and re-plans only the affected portions rather than starting over.
- At the engine level, both @aakashgupta and @testingcatalog noted the underlying architecture is an asynchronous Agent Team running in Deep Discover mode: the model decides for itself whether to split tasks, how many sub-agents to launch, and when to aggregate; sub-agents run in parallel, stream results back to shared task state, and feature adaptive recovery mechanisms.
- @testingcatalog mentioned that under the Agent Team configuration, Apodex 1.1 showed performance on APEX-A (benchmark), though specific figures were not detailed in the post.
Why it matters
- The release marks a clear shift in product positioning: from "answering questions" to "delivering verifiable work products," aligning with the industry's focus on agents' long-horizon task execution capabilities.
- Mid-task intervention and incremental re-planning address the real-world pain point of frequently changing requirements, a key engineering feature for making agents practical.
2026-09-16 ~ 2026-09-16 · 5 related posts
- Episode 1: Apodex launches 1.1, open-sources FrontierAgent framework and 35B mini model(2026-09-08, 14 posts)
- Episode 2: Apodex 1.1 Reorganizes Deep Research Pipelines on the Fly(2026-09-09, 3 posts)
- Episode 3: Apodex 1.1 Shifts from Research Q&A to Real Task Execution(2026-09-16, 5 posts)
- Episode 4: Apodex Open-Sources FrontierAgent Framework(2026-09-16, 2 posts)
- Episode 5: Apodex 1.1 Runs Real Survival Analysis with Reviewable Conclusions(2026-09-16, 2 posts)
Primary sources
- Apodex 1.1 launches: moves from deep research to real task execution — testingcatalog ·
- Apodex 1.1 argues a right answer can still be a failed task—execution is the unit of AI capability — aakashgupta ·
- Inside Apodex 1.1: Adaptive Recovery, AgentOS, and self-organizing async agent teams — testingcatalog ·
- [source] Apodex 1.1 launches: moves from deep research to real task execution — testingcatalog · 2026-09-16
- [source] Inside Apodex 1.1: Adaptive Recovery, AgentOS, and self-organizing async agent teams — testingcatalog · 2026-09-16
- [source] Apodex 1.1 argues a right answer can still be a failed task—execution is the unit of AI capability — aakashgupta · 2026-09-16
- Inside Apodex 1.1: an async Agent Team in Deep Discover mode on a live task board — aakashgupta · 2026-09-16
- Apodex 1.1 lets you drop new files mid-task, replanning only the affected parts — aakashgupta · 2026-09-16