35B agent approaches trillion-parameter performance by extending its horizon
A 35B Mixture-of-Experts agent reports 45K‑token trajectories and competitive long‑horizon scores against 1T‑parameter models; companion papers test self‑evolving world models and an interactive coding benchmark built from 11,260 sessions.
The business desk's take → Anthropic cuts cost of agentic AI with Claude Sonnet 530-second briefing
- 35B Agents-A1 matches 1T models via horizon scaling.
- Shifts leverage to scaffolding, distillation, and memory.
- Watch for public assets, compute reporting, and eval adoption.
One-Line Summary
Smaller agentic models push ahead by planning over longer sequences, while test-time world models and interactive coding benchmarks reshape how we evaluate agent performance.
Research Papers
35B agent approaches trillion-parameter performance by extending its horizon
This paper presents an AI agent that tackles complex, multi-step tasks by planning much farther ahead instead of simply adding more parameters. The team introduces Agents-A1, a35 billion-parameter Mixture of Experts ** (MoE) agent that targets long-horizon problems with very long multi-step trajectories. It reports average agentic trajectories of about45K tokens**, aiming to reach the results of trillion-parameter systems on long-horizon benchmarks. 1
To make long plans usable, the authors build a knowledge–action infrastructure that links external knowledge, actions, observations, and verifier outcomes into long trajectories. Training follows three stages: full-domain **supervised fine-tuning ** (SFT) to align base agent behaviors, domain-level teacher models to capture specialized expertise, and a multi-teacher, domain-routed on-policy distillation with salient vocabulary alignment that merges six heterogeneous domains into one deployable student. 1
On long-horizon agent benchmarks, Agents-A1 reports leading or competitive results versus trillion-parameter systems like Kimi‑K2.6 and DeepSeek‑V4‑pro: SEAL‑0 56.4,IFBench 80.6,HiPhO 46.4,FrontierScience‑Olympiad 79.0,MolBench‑Bind 56.8, with competitive scores onSciCode 44.3,HLE 47.6, andBrowseComp 75.5. The authors position horizon scaling as a practical path for a 35B agent to reach or match 1T-level performance on long-horizon tasks. 1
WorldEvolver: test-time memory helps agents plan better
WorldEvolver introduces a “world model” that updates what it remembers during deployment to make an agent’s lookahead more reliable—without changing the downstream agent or any model weights. In context, a **large language model ** (LLM) agent can use three modules: Episodic Memory (retrieval-based simulation from real transitions), Semantic Memory (heuristics extracted from prediction–observation mismatches), and Selective Foresight (filtering low-confidence predictions before they enter the reasoning context). 2
Evaluated on ALFWorld and ScienceWorld, the framework measures world-model prediction accuracy on Word2World and downstream agent success on AgentBoard. Experiments report the highest prediction accuracy across three backbones and stronger downstream success rates than other world-model baselines, indicating that test-time memory revision improves both foresight quality and planning outcomes. 2
SWE-Together: coding agents evaluated in real multi-turn sessions
SWE‑Together reconstructs real user–agent coding interactions so agents are judged as collaborators, not just by final code snapshots. The authors curate 109 repository-level tasks from11,260 recorded sessions, then replay them with a reactive**Large Language Model ** (LLM)–based user simulator that preserves original user intent and intervenes when the agent’s progress requires it. 3
The benchmark evaluates both final repository correctness and how many corrective feedback turns an agent needs. Results with frontier coding agents show a clear pattern: stronger agents tend to achieve higher final success while needing fewer interventions—evidence that multi-turn collaboration quality can be measured, not assumed. 3
Open Source & Repos
Goose: desktop and CLI agent that installs, runs, edits, and tests code
Goose is an extensible open-source agent with a desktop app, a command-line interface (CLI), and an application programming interface (API) that can install dependencies, execute programs, edit files, and run tests with any**Large Language Model ** (LLM)—released under the Apache‑2.0 license. The project has moved from block/goose to the Agentic AI Foundation at the Linux Foundation. 4
The latest release, **v1.39.0 ** (2026-06-25), highlights features for handling “recipes” and managing global configuration and session extensions—useful for packaging repeatable agent workflows. 4
Why It Matters
Long-horizon results from a 35B MoE agent suggest a shift from “more parameters” to “more horizon”—that is, investing in longer, verified trajectories and multi-domain distillation to teach agents how to plan and act over many steps. 1
At the same time, evaluation is moving toward reliability and interaction: test-time memory aims to make foresight trustworthy, and multi-turn coding benchmarks measure collaboration rather than single-shot accuracy. Together, these trends point to agent systems, memory, and measurement as the levers that matter most. 2
What to Try This Week
- Goose desktop/CLI: install from GitHub and run a starter “recipe” to automate a small code task. https://github.com/aaif-goose/goose
Terms in This Article
- AI Agent
An AI agent is a system that pursues a stated goal by perceiving its environment, reasoning and planning, and repeatedly acting through external tools or APIs, autonomously selecting the next action based on execution feedback and explicit stopping conditions.
- MoE
Mixture of Experts (MoE) is a sparse conditional-computation architecture in which a gating/router network selects a small subset of expert feed-forward subnets for each input token—typically replacing the Transformer’s FFN—so total parameter capacity can grow via many experts while per-token compute remains bounded.
- supervised fine-tuning
Supervised fine-tuning is the process of further training a pre-trained AI model using additional labeled data, where humans provide the correct answers, to adapt the model for specific tasks or goals. This step is essential for aligning the model's outputs with desired real-world results.
Two desks on this story
35B agent approaches trillion-parameter performance by extending its horizon
Anthropic cuts cost of agentic AI with Claude Sonnet 5
Anthropic makes Sonnet 5 its default consumer model and sets introductory API pricing at $2/$10 per million tokens through Aug 31, 2026. The pricing and capability mix pressures cost-per-task as Google trims media generation costs and Etched reports $1B in inference orders.
Comments (0)