Updates, announcements, trends and explainers from Stackness
One Playwright MCP snapshot cost my agent up to 33,917 tokens on 2 November 2026; a passing Playwright test cost 64. Look once, then write the test.
SESergei GordeichukSenior Software engineer · The master of StacknessCursor's Agent SDK went GA on 20 October 2026. Its prompts, skills and MCP configs port; its triggers, hosting and model calls stay on Cursor.
A loop of Qwen3.8-27B calls scored 92.4% to Fable 5's 90.4% on hard LiveCodeBench, not a significant gap, at $0.48 vs $0.61 a task and 15x the tokens.
Claude Code's status line uses 0 tokens. agent_summary took 2% to 24% of spend where measured. My startup context was 43,394 tokens.
Confirmed agent-stack key leaks in October 2026 ran through agent configs and hosted services, not .env. Where keys sit, how to test, and the caps.
Kimi K3.1 Pro costs $3.68 per Artificial Analysis task to Opus 5.5's $10.04 but scores 46 to 58.4. Which slot each takes, as of 30 October 2026.
Claude Code passed GitHub Copilot in the 2026 Stack Overflow survey, 68.4% to 61.3%. VS Code fell to 71.3% and Zed rose to 11.5%. Which moves are real.
Eight agent-built games were written up in October 2026, seven playable. Six run three.js in a browser, paid for by flat Claude plans, not per game.
ThinkRail 0.2.0 is an open source app that runs only the pi agent; JetBrains Air runs Claude Agent, Codex, Gemini CLI and Junie. Which to pick, as of October 2026.
Entire launched on 26 October 2026 with the Trail: agent evals score every push and gates decide the merge. What it replaces in a pull request, and what stays on GitHub.
One team's agent-written dashboards reached 156 weekly users against Looker's 157 by October 2026, at about $0.53 of tokens each. Looker stayed.
Agents built property-test suites for six open source repos in October 2026; by 28 October, 13 fix PRs had merged in uv, Prometheus and rv.
The Type Server Protocol lets Pylance in VS Code ask Pyright or Pyrefly for Python types as data. Version 0.4.1 changes the editor, not CI.
On SlopCodeBench 10/2026, xhigh cost 2.1 to 4.2 times medium. Opus 5.5 passed 28 of 196 checkpoints at medium and 25 at xhigh.
Google's Elephant Goldfish model drafts a spec in one long agent session, then has fresh-context subagents review it. v0.1.0 shipped on 27 October 2026.
When an agent's fix goes green and the bug stays, check first whether the new test can fail: run it on the parent commit, where it must fail.
A multi-user agent needs each call checked as the person it acts for, outside the agent's reach. A shared API key and a monitor in the repo fail that.
OpenBao fixed the chained Vault and OpenBao RCE 14 days before Vault. Pick OpenBao for Raft or PostgreSQL without Enterprise features; stay on Vault for them.
A 33-line architecture map loads 603 tokens per session, area notes load on demand, and a 21-line CI script fails when either names a deleted path.
Agent pull requests grow the Actions bill through repeat runs of one commit. Find them in the runs API first; cheaper runners list a minute at 33-79% less.
