Agent Tools Today, 7 Oct 2026: Mistral Large 4, EmbeddingGemma 2, OpenAI Decisions API

The top launches, skills and plugins, multi agent moves, rising repos and arguments in coding agents on 7 Oct 2026, each with its source.

Chaitanya Giri 5 min read News
Hand drawn sketch: a front page with five numbered lines, a clock, and a yellow character reading it out

Agent tools today, 7 Oct 2026, in one line: Mistral previewed its largest model, Google released a small embedding model, and OpenAI put its Decisions API into public beta. Every number was checked on 7 Oct, one day after the 6 Oct edition.

Animation. Five plant pots of different sizes land in a row on a shelf, numbered 1 to 5, from one very large pot to one very small pot, and a leafy plant grows in each. A paper tag swings down on a string under each pot: Mistral Large 4, EmbeddingGemma 2, OpenAI Decisions API, Claude Code v2.1.292 and Gemini CLI v0.63.0.
The five launches of the day, one pot per launch. Tap a pot for its one line.

Top 5 launches

  1. Mistral Large 4. Mistral opened a public preview on 6 Oct. Its post describes a multimodal model with 1 trillion parameters, 49 billion active, on 6 Oct figures. The preview API is live. The weights are not out: Mistral says it will release them by the end of October. Compare Reflection’s Beam. 1,625 points on Hacker News. Source
  2. EmbeddingGemma 2. Google released it on 6 Oct under Apache 2.0. One model with 740 million parameters, on Google’s 6 Oct figures, maps text, code, images, audio and video into one embedding space, sized to run on a device. 247 points on Hacker News. Source
  3. OpenAI Decisions API. In public beta, per a changelog entry dated 6 Oct. POST /v1/decisions returns typed answers: the probability that a condition is true, one choice from a fixed set, or a score against a rubric. gpt-6-luna is the only model so far. OpenAI says it is about 10x faster than the Responses API. 181 points on Hacker News. Source
  4. Claude Code v2.1.292. Released 6 Oct. It adds --marketplace <source> to claude plugin install and an effort parameter to the Agent tool. One fix is marked as security: it stops PreToolUse hook approvals and auto mode from bypassing the permission prompt for file reads from network (UNC) paths. Source
  5. Gemini CLI v0.63.0. Released 6 Oct. The notes list fixes: one bounds tool output size in long agent loops, one targets an endless sign in loop. Source

Top 4 skills, MCP servers and plugins

  1. html-plan. A plugin in Anthropic’s community marketplace for Claude Code. /html-plan turns a plan into one HTML page: a tree of claims you open one level at a time, with decisions you answer in place. Added 1 Oct, updated twice on 5 Oct. The marketplace repo is Apache 2.0, 4,527 stars on 7 Oct. Source
  2. OpenWork v0.18.57. Released 7 Oct. One prompt asks your agent to bring your Claude Cowork and Claude Code plugins and skills into OpenWork, a desktop app built on OpenCode. 23,920 stars on 7 Oct. Source
  3. pawl 0.4.1. One plugin for Claude Code, Codex and Antigravity that runs small checks as code before an agent uses a tool. The 6 Oct patch closes bypasses found in review: deletion checks now follow symlinks before deciding whether a command reaches a protected directory. Apache 2.0, 2 stars on 7 Oct, 2 points on Hacker News. Source
  4. skill-placebo v0.1.1. A test of nine popular skills against neutral text of the same length. The README reports that 2 of 9 beat the placebo, both on cost at an adjusted p of 0.049, 1 did worse and 6 did no better, across 450 trials with one model in Claude Code. The 6 Oct release recounts two timeouts as failures. MIT, 1 star on 7 Oct. Source

Top 5 moves from multi agent tools

  1. Orca. Release v1.4.222 on 7 Oct. A worker started with orca orchestration worker-start --agent opencode no longer leaves its task unsent in OpenCode’s input box on a busy machine: Orca now waits until OpenCode can submit it. Three other fixes live in the terminal background service and only apply once it restarts. Updating from v1.4.221 keeps the running service. Source
  2. Munder Difflin. Free and open source: no new release today. The latest is v0.5.5, from 1 Oct. Source
  3. Herdr GPUI. Release v20261006.1 on 6 Oct. It runs setup, run and archive scripts per repository for worktrees, and needs a deliberate click before it trusts them. Apache 2.0, 960 stars on 7 Oct. Source
  4. OpenCode. v1.18.35 on 6 Oct, the first release since 30 Sep. xAI tool results now include supported images. Source
  5. Pi. No new release since v1.0.4 on 5 Oct. One of its maintainers published “What is Codemode” on 6 Oct. In Pi, Codemode is JavaScript that runs in a sandbox on the harness side, so an agent can chain tool calls without each result passing through the model’s context. 25 points on Hacker News. Background: what is Pi agent. Source

Top 5 rising repos

  1. openai/math. Created 6 Oct. Mathematical manuscripts and proof files produced by an internal OpenAI model: 722 manuscripts in 372 families, per the README. Apache 2.0, 3,315 stars on 7 Oct. 72 points on Hacker News for one preprint. Source
  2. openTPU. An open source AI accelerator, developed by AI, per its README. The hardware design, instruction set, simulator and compiler sit in one repo. Apache 2.0, 248 stars on 7 Oct, 251 points on Hacker News. Source
  3. OpenChart. A desktop charting app with AI agents for market analysis and alerts. Created 5 Oct. Modified Apache 2.0 licence, 92 stars on 7 Oct, 38 points on Hacker News. Source
  4. Jiti. Grow a running Common Lisp application through chat. Created 4 Oct. No licence listed. 53 stars on 7 Oct, 29 points on Hacker News. Source
  5. AnyPS5. Not an agent tool. It ports PS5 executables to Linux and Windows without emulation. GPL 2.0, 6,769 stars on 7 Oct, 134 points on Hacker News. Source

New repos can gain stars quickly for reasons other than use, so treat these as names to watch, not recommendations.

Top 3 things people are arguing about

  1. Should a model refuse security work? Mistral’s 6 Oct post says several closed models, Claude Opus 5.5 and GPT-6 Astra among them, score near zero on one vulnerability test because they refuse the task, and that Mistral Large 4 scores 82% on it. 974 comments on Hacker News. Source
  2. Is a proof that nobody checked a result? The openai/math README says not all manuscripts have Lean formalizations, and that some of the unformalized results could have issues. 534 points and 455 comments on Hacker News. Source
  3. Is vibecoding less fun? An essay from 20 Aug 2026 reached Hacker News on 6 Oct. It argues that vibecoding puts the fun at the start, and that refactoring the result is not thrilling. 180 points and 258 comments. Source

How this list is built

Each day our agents read GitHub releases and trending pages, the Hacker News front page, and the top posts on X and Reddit about coding agents. A name only makes a list with a link that was opened and a signal that can be shown: points, stars or a dated release. Repos are checked against the GitHub API before they appear, and a claim that exists only as a social post is left out. New here? Start with the install guide.