Sydney Runkle reported that integrating a model router into a coding agent's harness cut operational costs by 64% without reducing output quality. The four-step workflow involves understanding tasks and models, embedding the router in the harness, and tracking task outcomes.
A UT Austin study ran nearly 35,000 coding-agent runs on SWE-bench Verified and Terminal-Bench while varying how context is compressed, when it triggers, and how much is removed. On Terminal-Bench with Qwen, policies using about a third of the tokens can take 20% to 80% longer than keeping full context.
In the GitHub Copilot app, the diff, terminal, and browser sit together so reviewers can check, run, and preview agent code without switching tabs. A beginner guide covers using that layout.
AutoCompact uses a judge to correct compaction decisions, then applies supervised fine-tuning and reinforcement learning with task-success rewards. Reported pass rates rise by 9.2 points on SWE-bench Verified and 5.0 points on SWE-PolyBench Verified, including when a 256K window never overflows.
OpenAI Developers says Codex cloud environments let agents keep working after a laptop is closed. Reusable environments come with the repository, dependencies, scripts, and settings already in place.
OpenRouter had claude-opus-5.5 grade every run of its openrouter-decisions skill against a fixed checklist. In a blind comparison, Opus picked the skill version of each task all 36 times.
OpenRouter’s new openrouter-decisions skill points a coding agent at a classification, routing, or moderation step and rebuilds that step on a decision model through the Decisions API.
Melissa Pan compared Claude Code, Codex, and Pi and reports that the system around an AI model, the harness, affects cost more than accuracy. She cites three findings; only this one is stated in the announcement.
Google Gemma and Kaggle have opened the Gemma 4 Developer Agent Competition, offering a prize pool of more than $110,000. Entries close on November 2, 2026, and a starter kit is available on the competition page.
In the GitHub Copilot app, agents can run in parallel, with each session using its own Git worktree and context so building, reviewing, and testing can happen at the same time.
CASD has an off-the-shelf coding agent analyze a full set of agent logs and write the findings as rules in one prompt, with no environment access or validation data. DAIR.AI says one pass beat GEPA and SkillOpt on the reported benchmarks at about $1.60 per optimized prompt.
Administrators can manage models, tools, policies and budgets in Unity Gateway while developers keep using Claude Code, Codex, OpenCode and other coding agents. One configuration update is applied across supported agents.
At max effort in Claude Code, Artificial Analysis measures Opus 5.5 at 66, above Opus 5 at 60 and Claude Fable 5.1 at 62. Its cost per task is $13.04, higher than Opus 5’s $10.79 despite lower token prices.
Token Meter is a free, open-source menu bar tool from Splunk that shows coding agent costs as a run is underway.
The short-lived environments capture both schema and data, so tests do not affect production. Workflows use the Databricks CLI and AGENTS.md to compare changes with the parent branch before approved updates move through CI/CD.
Databricks says DevHub gives developers ready-to-use templates, documentation, step-by-step guidance, and agent prompts that can be copied into Cursor, Claude Code, Codex, or another coding agent.
elvis (@omarsar0) says he early-tested StepFun’s new Step 5 Preview as a coding agent. He reports that it and GLM 5.3 both solved two real repository tasks with no fix rounds, that Step 5 Preview stopped after checking its work, and that it found five hidden clues in a ~368K-token long-context test.
