· 3 min read
The Gradient — 2026-10-10
- Company Blog
- Research
In a striking efficiency boost, Asana leveraged GPT‑6 Astra in Codex to dramatically cut the cost and speed up its browser‑based AI agent.
Asana cuts model costs 76x in browser tests with GPT-6.1 Sol
What happened: Asana integrated the GPT‑6.1 Sol model (referred to as GPT‑6 Astra in Codex) into its browser agent, achieving a 76‑fold reduction in compute cost and a 5‑fold increase in speed during internal tests.
Why it matters: The cost savings make high‑performance AI agents more affordable for customers, paving the way for broader adoption of sophisticated browser‑based automation and productivity tools.
Key stats:
- Cost reduction: 76× cheaper
- Performance gain: 5× faster
- Model used: GPT‑6.1 Sol (GPT‑6 Astra)
The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate
What happened: Researchers argue that self‑evolving LLM agents must keep their model weights fixed and confine all adaptations to the runtime harness (instruction text, tool‑call logic, and primitive composition), generating a traceable diff with a cause and a test attached.
Why it matters: This restriction preserves auditability and regulatory compliance, ensuring supervisors can review every change without losing the original artefacts (named change, test, approval) that underpin credit decisions.
Key stats:
- Introduces a "compliance‑bounded" self‑evolution framework.
- No empirical results yet; the contribution is conceptual and architectural.
- Guarantees 100 % weight immutability while allowing unlimited harness modifications.
Source: https://arxiv.org/abs/2610.10629
Plan-and-Patch: Diffusion Language Models for Agentic Planning
What happened: The authors present Plan‑and‑Patch, a diffusion language model that first drafts a multi‑step plan and then uses a diffusion‑driven “patch” step to edit only the portions that become invalid due to environmental changes, tool failures, or unexpected outcomes.
Why it matters: By allowing selective revisions instead of discarding entire plans, agents become more efficient and robust, handling real‑world uncertainties while reducing computation and error propagation.
Key stats: Evaluated on several long‑horizon benchmark suites, Plan‑and‑Patch achieved higher task‑completion rates than static planning baselines, demonstrating superior adaptability and fewer total actions needed for successful execution.
Source: https://arxiv.org/abs/2610.10786
On the Clock: Towards Punctual and Productive Time-Budgeted AI Agents
What happened: Researchers tested small language‑model agents—Qwen3.6‑27B on five MLE‑Bench Lite contests and Qwen3‑4B on the Zork I (Jericho) text‑adventure—under explicit wall‑clock budgets. The budget was communicated only via the prompt, and agents were required both to stay within the allotted runtime and to use any spare seconds productively.
Why it matters: Demonstrating that AI systems can self‑regulate time opens the door to more reliable, cost‑effective deployment in real‑world settings where compute budgets are tight (e.g., edge devices, API billing). It also suggests a new lever—time management—for improving agent performance without altering model size.
Key stats:
- Qwen3.6‑27B achieved measurable score improvements across all five MLE‑Bench Lite tasks when given modest extra seconds.
- Qwen3‑4B’s performance on Zork I rose noticeably, confirming that additional compute time can translate into deeper exploration and better decision‑making.
- The agents consistently honored the prompt‑specified time caps, evidencing emergent time‑budget awareness.
Source: https://arxiv.org/abs/2610.10833
StoreBench: A Live-Commerce Environment for Evaluating and Training Autonomous Operator Agents
What happened: Researchers unveiled StoreBench, a live‑commerce simulation where autonomous agents manage a mid‑size online clothing store using a production‑grade backend. The environment runs continuously, exposing agents to realistic customer behavior, inventory changes, and revenue pressures.
Why it matters: Traditional RL benchmarks freeze the world until the agent acts and reward only at episode end, limiting long‑term planning. StoreBench introduces a dynamic, ongoing marketplace, enabling the training and evaluation of agents that must balance inventory, pricing, marketing, and fulfillment over weeks of simulated operation.
Key stats: • Simulates a store with thousands of SKUs and realistic inventory levels
• Handles tens of thousands of daily simulated customer visits
• Supports episodes up to 30 days of continuous operation
• Integrated with a production‑grade e‑commerce stack (catalog, checkout, analytics)
Source: https://arxiv.org/abs/2610.10942