Apodex 1.1: Asynchronous Agent Team Arrives, Plus Open-Source FrontierAgent

August 24, 2026 — Apodex shipped Apodex 1.1, two months after Apodex-1.0-mini topped the FutureX leaderboard running on open 35B weights. The headline change is not raw benchmark lift — it's where the multi-agent behavior lives. Apodex 1.1's asynchronous Agent Team is a trained capability inside the model, not an orchestration script wrapped around it, and it comes with an open-source local workbench, FrontierAgent, so anyone can run the same ReAct-vs-Agent-Team comparison on their own machine.

TL;DR

What Agent Team actually is

Most "multi-agent" products today are a harness problem: a supervisor script calls a model N times, merges outputs, and calls that orchestration. Apodex is arguing that decomposition itself should be a trained skill.

In Apodex 1.1, the model autonomously decides three things at inference time:

1. Whether to decompose a task into sub-agents at all, versus running it as a single agent
2. How many sub-agents to deploy for the task at hand
3. When to consolidate results back into a single deliverable

Sub-agents explore branches in parallel and feed findings into shared task state continuously — the model doesn't block on the slowest branch before it starts synthesizing. Apodex describes this as a "continuous feedback loop between the model, the Subagents, and the user," and it's built to support mid-task intervention: a user can interrupt, redirect, or add a requirement without discarding work already completed by other branches.

That's a similar problem shape to Anthropic's managed multi-agent orchestration work and the patterns catalogued in explainx.ai's multi-agent orchestration guide — the industry-wide question of whether coordination should live in the harness or in the model. Apodex's bet is squarely on the model.

The measured lift

Apodex evaluated four benchmarks in both a standard ReAct configuration and Agent Team configuration on the same underlying model. Agent Team added 4.1 to 9.3 points across those four — a meaningful, if not dramatic, gain purely from letting the model choose its own decomposition strategy instead of running single-threaded.

Benchmark scores — Apodex 1.1 (full model)

FrontierFinance and FrontierScience-Research are Apodex's own agentic benchmarks, aimed at the "complex professional work" niche the model targets — financial modeling with tool use, and scientific research requiring multi-step evidence synthesis. GDPVal is scored as a win-rate against a reference response rather than a raw accuracy number, which is worth flagging when you compare it to accuracy-style benchmarks elsewhere on this table.

Apodex 1.1 mini — the open-weight story

Apodex 1.1 mini is a 35B parameter model, released with open weights specifically for local deployment. The comparison Apodex leads with is deliberate: 35B parameters, running in Agent Team mode, landing within a point of a model roughly 28x its size.

Two things stand out here. First, Agent Team mode is doing real work at the mini scale too — it's not just a full-model feature, adding roughly 6-10 points across all three benchmarks over plain ReAct. Second, "virtually matching" a 1T-parameter model on APEX-Agents at 35B parameters is the kind of claim worth treating the way explainx.ai treated Apodex's FutureX numbers in June: a strong signal on Apodex's own benchmark suite, not yet an independently reproduced result. Run it against AgentHarness or your own eval set before betting a procurement decision on it, the same caution that applied to the Apodex-1.0-mini FutureX claims.

For context on why a 35B open model beating a trillion-parameter closed one is plausible rather than implausible, see explainx.ai's breakdown of what parameter counts actually tell you — training data quality and agentic RL recipes routinely outweigh raw scale on task-specific evals, which is the same story Agents-A1's 35B MoE model told in June.

FrontierAgent — the open-source workbench

Alongside the weights, Apodex open-sourced FrontierAgent: a locally deployable research workbench built for the Apodex 1.1 family.

Shipping ReAct and Agent Team as a toggle in the same open harness is the useful part for builders: you can point FrontierAgent at Apodex 1.1 mini locally, run the same task both ways, and see the coordination overhead versus the coordination lift for your own workload — rather than trusting Apodex's benchmark table on faith. That's the same instinct behind explainx.ai's agent harness guide: the harness determines what a model can actually do in production far more than the raw checkpoint does.

For teams already running local agent stacks, FrontierAgent slots into the same category as running open models locally with OpenCode or Kimi K3 on your own hardware — one more no-Docker, single-binary path to a fully local agent loop, which matters for the closed vs. local open-source tradeoff teams are increasingly forced to make.

Apodex 1.1 vs Apodex 1.0-mini — what actually changed

The throughline across both releases is Apodex betting that architecture and training recipe beat parameter count on agentic, evidence-heavy tasks — first via verification-centric training in 1.0, now via trained-in task decomposition in 1.1.

Skepticism to keep

Bottom line

Apodex 1.1's real contribution is putting multi-agent coordination inside the model's training loop and then proving the delta with an apples-to-apples ReAct-vs-Agent-Team comparison on the same weights — a cleaner experiment than most "agent team" product launches offer. Combined with a 35B open mini that's competitive with a trillion-parameter closed model on at least one benchmark, and an open, no-Docker CLI harness to test it yourself, Apodex 1.1 is worth a spot in your local-agent evaluation queue even before platform.apodex.ai's API access fully rolls out.

If you're evaluating it, run three checks:

1. FrontierAgent locally in both modes on your own task, not just Apodex's benchmark suite
2. Apodex 1.1 mini against your current local model of choice — see the AI benchmarks guide for how to read the comparison honestly
3. Cost and access tiering against closed-source vs. local open-source alternatives before committing a production workload

Related on explainx.ai

• Apodex 1.0-mini — FutureX #1, 35B deep research — the predecessor release this builds on
• Multi-agent orchestration patterns guide — where Agent Team fits among coordination approaches
• What is an agent harness — complete guide — why FrontierAgent's ReAct/Agent Team toggle matters
• Anthropic managed agents and multi-agent orchestration — how another lab is approaching the same coordination problem
• Agents-A1 — InternScience 35B MoE — another 35B open agent model from the same weight class
• Closed-source AI vs. local open-source alternatives — the tiering decision Apodex 1.1 mini feeds into
• Kimi K3 — run locally on your own hardware — comparable local-deployment path
• AI benchmarks complete guide — how to read APEX-Agent, GDPVal, HLE, and vendor tables generally

Official: apodex.ai · platform.apodex.ai · Hugging Face — Apodex 1.1 collection · FrontierAgent on GitHub

Benchmark scores and feature details reflect Apodex's August 24, 2026 announcement and technical paper. Model access, API rollout status, and leaderboard positions change quickly — verify current availability at apodex.ai and platform.apodex.ai before citing figures. Last updated: August 25, 2026.