#coding-agents

Every published rate of coding agents gaming their tests, with the count or population it is over: METR, ImpossibleBench and a September 2026 paper. 19 rows.

A UC Berkeley and Arena study ran seven models in three coding harnesses. Success barely moved; cost moved up to 5x. All 42 measured rows, with intervals.

Anthropic says coding agents raised its CI jobs 25x in six months. Three patches bought 70 days, 29 days and under a day. What the redesign teaches.

Check Kimi K2.8 Preview access, context limits and reasoning settings, then plan a coding trial that separates official specifications from performance claims.

Assess Cognition SWE-2 for coding work. Check Devin access, the limited free-use offer, benchmark scope and review costs before choosing a paid plan.

Test whether an AI agent uncovers missing client requirements before building. Compare interview quality, code inspection, evidence and acceptance criteria.

Plan an Astra-assisted game around a playable loop, repeatable checks and human feedback. Use a practical playtest method before investing in more polish.

Terminal-Bench 4.0 changes the test behind agent scores. Decide when results need rerunning, regrading or reuse before comparing coding agents for your team.

Test AI-built forms beyond a successful submit. Preserve valid input, explain errors and distinguish a rejected request from an outcome still unknown.

Scope a small AI-built utility around clear inputs, outputs and limits. Decide what it should own, reject and preserve before it grows into a system.

Give a coding agent reproducible bug evidence: exact steps, expected behavior, safe sample data and a failure it can observe before proposing a fix.

OpenAI, Cloudflare, Ramp and Google Chrome published their own numbers from agents finding and fixing security bugs. One table, every definition, every gap.

A census of what eleven coding-agent harnesses isolate by default and by flag: filesystem scope, network egress, process spawning, credential visibility.

Default-allow's real cost is not a model mistake but a documented attack primitive. When ask-first is theatre, and why reversibility should route the call.

Eighteen AI coding tools, the instruction files each reads, precedence when several apply, and load behavior — from 27 vendor docs, retrieved Aug 30, 2026.

The tool with the widest reach over a marketer's week is a coding agent. Which data and file tasks to hand it, and how to check work it reports as done.

Headless permission defaults for 12 coding-agent CLIs: which write files without asking, which refuse until you pass a flag, and which actually sandbox.

One fixed Python task, run twice through eight headless coding-agent CLIs, with tokens, wall-clock time and list-price cost measured for every run.

Claude Code v2.1.234 hardened the remaining pre-approval NTLM path accesses. Codex CLI 0.148.0 added Bedrock and session forking. What changed for operators.

CVE-2026-75130 scores 9.0 under CVSS 3.1 and 6.4 under CVSS 4.0. Context7 2.1.2 and earlier are affected; no public fix documented as of Aug 22, 2026.

Seventeen coding agents, six questions each, vendor documentation only. What the published terms say on retention and training, and which cells stayed open.

Grok 4.6 reached Cursor the day it launched. The argument that switching cost has moved to the harness, set out with its counter-evidence alongside it.

DeepSeek open-sourced its dsh agent harness under MIT: a developer preview at 0.1.0-rc.5, four runtime modes, and every part of the product a plugin.

GLM-5.3 reuses GLM-5.2's base model, so Z.ai says every gain is post-training. Terminal-Bench 3.0 jumped 4.6 to 28.3, and weights are promised later.

Both terminal harnesses ship bash sandboxing off by default. The differences that hold up are openness, model freedom and what each vendor publishes.

Meta shipped Muse Code beta and Muse Spark 1.2 together on August 5, at two prices: $1.25/$4.25 standard, or $0.10/$0.20 if you contribute data.

Muse Code fans work out to a git worktree per subagent — opt-in, in git repos — records every step in an append-only event log, and ships four built-in skills.

poolside trains across several agent harnesses on purpose, Cursor stitches tool-call corrections into training, Anthropic argues for holding the model fixed.

OSWorld measures desktop GUI control, Terminal-Bench measures CLI work. Same-named benchmarks differ by version and harness, so check before you compare.

Claude Code shipped isolation fixes in four of five releases. Separately, GuardFall research defeated command filters in 10 of 11 agents.

OpenAI's $230 Codex Micro keypad solves approval latency across parallel agent runs, not typing. Why supervising agent fleets is now the real UX bottleneck.

Hunt.io found Claude Code and DeepSeek-v4-pro wired into a suspected China-linked intrusion. The agent-governance controls that close the gap it exploited.

Grok Build reportedly uploaded full repos and git history to xAI's cloud. The /privacy toggle never stopped it — a server-side flag did. Agent trust is infra.

Windsurf silently became Devin Desktop on June 2, 2026. Your plans and settings carried over, but Cascade is EOL July 1. What every user should do now.

2026 benchmark comparison of OpenClaw, Hermes Agent, and Codex CLI. OpenRouter token data, real performance numbers, and the decision matrix.

OpenAI's Codex Subagents reaches GA with manager agents coordinating specialized workers across repositories. Setup guide for multi-agent coding workflows.

OpenAI launches Codex as a native Windows app with open-source agent sandbox, parallel tasks, PowerShell integration, and per-task Git worktrees. Setup guide.

Cursor launches Automations: always-on coding agents triggered by Slack, Linear, GitHub, and webhooks. Guide to event-driven autonomous coding.

Windsurf replaces credits with quota-based pricing at $20, $40, and $200 tiers. Comparison with Cursor and Copilot pricing for AI coding tool budget planning.

Windsurf Wave 13 adds Arena Mode for blind model comparison, Plan Mode for smarter task planning, parallel multi-agent sessions, and quota-based pricing.

Windsurf SWE-1.5 delivers 950 tok/s speed—14x faster than Claude. RLHF training, Cerebras WSE infrastructure, and real agency applications.

GitHub Agent HQ unifies AI coding agents from Anthropic, OpenAI, Google, and more. Mission Control, enterprise governance, and MCP integration explained.

Cursor 2.0's Composer model delivers frontier intelligence 4x faster. Multi-agent workflows, RL training, browser testing, and enterprise security explained.