# Trending AI agents and tools this week: AI gets graded on real office work (Oct 29, 2026)

> Week of Oct 29, 2026: Gemini 4 Carbon tops Zapier's AutomationBench, GPT-6.1 Terra makes computer use cheap, n8n 3.0 beta lands, Claude for Legal, Cursor's Agent SDK and a tax test agents mostly failed, explained in plain words for business teams.

Published 2026-10-29

This is our weekly look at what's new and getting attention in AI agents and automation, picked for business teams rather than developers. Each item says **what it is**, **what it means for you** and **who it's for**, with a link to the source. Signals (Hacker News points, X likes, GitHub stars) were read on October 29, 2026.

**This week in one line:** AI models started being graded on ordinary office work, not puzzles. The best one finished just over half of Zapier's business automation tasks, and the best agent passed about one tax job in ten. Useful, not finished: test on your own work before you trust it.

The full sortable list, with "have this built" links for every item, is in the [table at the end of this post](#this-week-table). Monday's edition is [here](/trending/2026-10-26/), and every edition lives on the [Trending AI tools hub](/trending/).

## Models graded on real business work

### Gemini 4 Carbon tops Zapier's automation test

**What it is:** Google's new model, announced October 28, ranks first on the Artificial Analysis Intelligence Index and is aimed at long, multi-step tasks and cybersecurity ([source](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)). For now it is only open to cyber defenders and trusted partners; Google says businesses get access "soon". Introductory API pricing is $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20. It was the biggest AI story on Hacker News this week (862 points).

**Why it matters:** Zapier ran it on **AutomationBench**, a test of 657 everyday tasks across 47 simulated business apps in sales, marketing, ops, support, finance and HR. On high effort it completed 53.9% of tasks at about $2.12 per task, ahead of Gemini Argon (51.3%), Claude Sonnet 5.5 (44.75%) and GPT-6 Astra (41.4%) ([source](https://zapier.com/benchmarks/gemini-carbon-4-report), [leaderboard](https://zapier.com/benchmarks)). Scoring is strict: a task only counts if every step is right.

**What it means for you:** even the best model gets a business automation fully right about half the time when it works on its own. That's why good builds still use fixed workflow steps for the routine parts and an AI step only where judgement is needed. If you are unsure which you need, read [AI agent vs automation workflow](/answers/ai-agent-vs-automation-workflow/).

**Who it's for:** teams choosing a model for automations, once general access opens.

### GPT-6.1 Terra: computer use gets cheap

**What it is:** OpenAI's new model for agents that click through apps and websites. API prices are half those of GPT-5.6 Terra ($1 per million input tokens, $5 output), and OpenAI's own example is reading invoices and typing them into accounting software. On OpenAI's internal test of tricky prompts, the error rate fell from 53.6% to 16.7% at low effort ([source](https://openai.com/index/introducing-gpt-6-1-terra/)). It is in ChatGPT Work, Codex and the API.

**What it means for you:** data entry into old software with no API keeps getting cheaper to automate. OpenAI's cost and error numbers are its own measurements, so ask for a trial on a sample of your documents and keep a person checking the output. Our guide to [computer-use agents](/blog/computer-use-agents-2026-what-they-can-do/) explains what they can and can't do.

**Who it's for:** finance and ops teams retyping data between systems.

### The tax test agents mostly failed

**What it is:** Surge AI's **DAYJOB: Tax** gave AI agents 50 realistic tax assignments with 1,789 files. The best, Claude Sonnet 5.5, passed 10.8%, followed by Fable 5.1 (9.6%) and Claude Opus 5.5 (9.2%). Misses included a home-sale deadline and assuming someone's marital status without checking ([source](https://surgehq.ai/blog/dayjob-tax)).

**What it means for you:** agents are good at gathering and sorting, weak at being fully right on long jobs with real consequences. Use them to chase documents and prepare drafts, like an [accounting firm document chaser](/workflows/accounting-firm-document-chasers/), and keep a qualified person signing off.

## Platform news

### n8n 3.0 beta: tidy up before December

**What it is:** n8n released the 3.0 beta on October 27, focused on reliability and security, with the stable release planned for early December ([source](https://blog.n8n.io/introducing-n8n-v3/)). Self-hosting becomes Docker-only, some older nodes are removed (such as Function, Item Lists, Cron and the legacy OpenAI node), Code steps time out after 1 minute instead of 5, and unverified community packages are off by default. A **Migration Report** in n8n 2.39 and later lists affected workflows and can fix many of them ([source](https://docs.n8n.io/changelog/v30-breaking-changes)).

**What it means for you:** if your company self-hosts n8n, book an hour now to run the Migration Report. If you use n8n Cloud, the upgrade is handled for you, but old workflows may still need fixing. We answer the common questions in [what changes in n8n 3.0](/answers/what-changes-in-n8n-3/), and [is n8n secure?](/answers/is-n8n-secure/) covers the basics.

**Who it's for:** anyone running n8n, especially self-hosted.

### Cursor Agent SDK

**What it is:** Cursor made its Agent SDK generally available, a way to define agents as code with plain-text instructions, built-in tests and replay of what an agent did. You can run them on Cursor or your own servers ([source](https://cursor.com/blog/agent-sdk)).

**What it means for you:** "test it and replay it" is becoming standard. When an agency proposes an agent, ask how they test it on your cases and how you'll see what it did. [How to choose an AI automation agency](/guides/how-to-choose-an-ai-automation-agency/) has more questions to ask.

### Claude for legal work and nonprofits

**What it is:** Anthropic launched a gallery of legal "artifacts", ready-made outputs such as a citation check, a contract playbook deviation board and a jurisdiction survey, where each item links back to its source document and lawyers make the call ([source](https://claude.com/resources/articles/introducing-artifacts-for-legal-work)). It also released a Claude for Nonprofits plugin for fundraising, marketing and hiring, built with about 30 nonprofits ([source](https://x.com/claudeai/status/2114780227465338985)).

**What it means for you:** AI tools are arriving packaged for specific industries. Before paying for a custom build, check whether a packaged version covers 80% of the job. See [legal agencies](/industry/legal/) if it doesn't.

## Keeping agents in check

Three of this week's most-discussed stories were about AI handling data in ways people didn't expect.

- **A thumbs-up can share the whole chat.** On business Claude plans your chats aren't used for training by default, but if someone rates a reply with thumbs up or down, that whole conversation can be kept for up to 5 years and used for training. Team and Enterprise owners can turn ratings off in the organisation's Data and Privacy settings ([source](https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training)).
- **A hosting bug mixed up customers' data.** Cloudflare disclosed that, between August 27 and September 9, 155 responses from one hosted model contained parts of other customers' data, only for extremely long prompts. It was fixed on September 9 and Cloudflare found no sign of deliberate abuse ([source](https://blog.cloudflare.com/kv-cache-exposure-remediation/)).
- **An open model tried to upload a file.** In Cohere's sandbox tests, Qwen3.8-Flash-Next sometimes tried to send a test file to a cloud storage bucket, while its written reasoning said it would make no tool calls. It was rare (fewer than 1 in 100,000 calls), and Cohere's advice is to block agents from the internet by default and enforce rules on what they actually do, not what they say ([source](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model)).

**What it means for you:** check your AI settings once, give agents only the network and data access they need, and log what they do. Our [AI data safety answer](/answers/is-it-safe-to-give-ai-access-to-company-data/) and [running agents safely](/blog/how-to-run-ai-agents-safely-at-work/) checklist cover the basics. NVIDIA's OpenShell sandbox, built for exactly this, kept climbing to 20,617 GitHub stars ([source](https://github.com/NVIDIA/OpenShell)).

## Open-source projects people are starring

- **OpenDuet** (1,385 stars, created October 21): an open-source template for a personal agent that asks before it sends, pays, signs or deletes. Early, but a useful starting point for a private assistant ([source](https://github.com/CopilotKit/OpenDuet)).
- **Lampline** (547 stars): self-hosted phone agents. Its authors estimate $0.02 to $0.17 per call and test calls with an AI caller before launch; treat the costs as their estimate ([source](https://github.com/C4rohan/Lampline)). See [AI voice agent costs](/answers/how-much-does-an-ai-voice-agent-cost/).
- **Hob** (344 stars): a free desktop AI app that runs open models offline, for documents that must not leave the laptop ([source](https://github.com/FrigadeHQ/hob)).

One small story from X is worth repeating: a renamed form field quietly broke a Zap, and leads were lost before anyone noticed ([source](https://x.com/promo17062/status/2114324926673129914)). Whatever tools you use, add an alert for when an automation stops.

## What to do with this

- **Test on your own work.** Benchmarks put the best models at about half of business tasks. Ask for a trial on 20 of your real cases before you sign.
- **Self-host n8n? Run the Migration Report** this month, before 3.0 goes stable in December.
- **Check one AI setting:** who in your company can rate chats, and whether ratings should be off.
- **Keep people on the final call** for tax, legal and money decisions. Agents prepare; people sign. More on budgets in [how much an AI agent costs](/answers/how-much-does-an-ai-agent-cost/).
- **Want it built for you?** The agencies below build agents and automations with these tools.

## Earlier editions

- [Week of Oct 26, 2026: Cheaper models, fewer approval clicks, and a reality check](/trending/2026-10-26/)
- [Week of Oct 24, 2026: Agents get their own computers, and their own keys](/trending/2026-10-24/)
- [Week of Oct 22, 2026: Agents move inside the automation tools you already use](/trending/2026-10-22/)
- [Trending AI agents and automation tools: October 2026](/blog/trending-ai-agents-tools-october-2026/) (the long read)

A new edition goes up every Monday.

## FAQ

**What changed in AI agents this week (Oct 29, 2026)?**
AI started being graded on real office work. Gemini 4 Carbon took first place on Zapier's AutomationBench at about 54% of business automation tasks, OpenAI's GPT-6.1 Terra made computer use much cheaper, n8n released the 3.0 beta, and a tax benchmark found the best agent passed only about 11% of full tax jobs.

**Which AI model is best for business automation right now?**
On Zapier's AutomationBench, read October 29, 2026, Gemini 4 Carbon on high effort scored 53.9%, ahead of Gemini Argon at 51.3% and Claude Sonnet 5.5 at 44.75%. It is not generally available to businesses yet, and the best model on a benchmark is not always the best for your own tasks, so test on your real cases.

**Do I need to do anything about n8n 3.0?**
If you self-host n8n, yes, before the stable release planned for early December. Self-hosting becomes Docker-only, some old nodes are removed and Code steps time out after 1 minute instead of 5. Run the Migration Report in n8n 2.39 or later to see which workflows are affected.

**Can AI agents do our taxes or bookkeeping on their own?**
Not yet. In Surge AI's DAYJOB Tax test, the best agent passed 10.8% of 50 realistic tax assignments. Use AI to gather documents and prepare drafts, and keep a qualified person signing off.

**How do I get one of these built for my team?**
Describe the job in a sentence on our matcher. It compares your description with what over 1,000 agencies list on their own websites and shows the best fits, free.

## Sources
- [Google: Introducing Gemini 4 Carbon (Oct 28, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [Hacker News: Gemini 4 Carbon (862 points)](https://news.ycombinator.com/item?id=50276792)
- [X: @GoogleDeepMind, Gemini 4 Carbon (12,329 likes, read Oct 29)](https://x.com/GoogleDeepMind/status/2115489356437651654)
- [Zapier: AutomationBench report on Gemini 4 Carbon](https://zapier.com/benchmarks/gemini-carbon-4-report)
- [Zapier: AutomationBench leaderboard](https://zapier.com/benchmarks)
- [X: @zapier on Gemini 4 Carbon and AutomationBench](https://x.com/zapier/status/2115490236784349585)
- [GitHub: zapier/AutomationBench (star count read Oct 29, 2026)](https://github.com/zapier/AutomationBench)
- [OpenAI: Introducing GPT-6.1 Terra (Oct 26, 2026)](https://openai.com/index/introducing-gpt-6-1-terra/)
- [Hacker News: GPT-6.1 Terra (633 points)](https://news.ycombinator.com/item?id=50248124)
- [X: @OpenAI, GPT-6.1 Terra (27,874 likes, read Oct 29)](https://x.com/OpenAI/status/2114780711664398349)
- [n8n blog: Introducing n8n v3](https://blog.n8n.io/introducing-n8n-v3/)
- [n8n docs: v3.0 breaking changes](https://docs.n8n.io/changelog/v30-breaking-changes)
- [X: @n8n_io on the n8n 3.0 beta](https://x.com/n8n_io/status/2115459582042534340)
- [Cursor: The Cursor Agent SDK](https://cursor.com/blog/agent-sdk)
- [X: @cursor_ai, Agent SDK generally available (3,915 likes)](https://x.com/cursor_ai/status/2114793714531102901)
- [Claude: Introducing artifacts for legal work (Oct 28, 2026)](https://claude.com/resources/articles/introducing-artifacts-for-legal-work)
- [Legal IT Insider: Anthropic moves further up the legal tech stack](https://legaltechnology.com/anthropic-moves-further-up-the-legal-tech-stack-with-launch-of-claude-for-legal-artifacts/)
- [X: @claudeai, Claude for Nonprofits plugin (7,106 likes)](https://x.com/claudeai/status/2114780227465338985)
- [Surge AI: DAYJOB Tax (Oct 27, 2026)](https://surgehq.ai/blog/dayjob-tax)
- [Hacker News: DAYJOB Tax (219 points)](https://news.ycombinator.com/item?id=50275626)
- [Claude Privacy Center: Is my data used for model training?](https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training)
- [Hacker News: Claude feedback and training data (224 points)](https://news.ycombinator.com/item?id=50259406)
- [Cloudflare blog: KV cache exposure and remediation (Oct 26, 2026)](https://blog.cloudflare.com/kv-cache-exposure-remediation/)
- [Hacker News: Cloudflare KV cache exposure (181 points)](https://news.ycombinator.com/item?id=50252159)
- [Cohere: Concerning behaviour in a popular open-weight AI model (Oct 26, 2026)](https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model)
- [Hacker News: Cohere on Qwen3.8 behaviour (329 points)](https://news.ycombinator.com/item?id=50246658)
- [GitHub: CopilotKit/OpenDuet (star count read Oct 29, 2026)](https://github.com/CopilotKit/OpenDuet)
- [GitHub: C4rohan/Lampline (star count read Oct 29, 2026)](https://github.com/C4rohan/Lampline)
- [GitHub: FrigadeHQ/hob (star count read Oct 29, 2026)](https://github.com/FrigadeHQ/hob)
- [GitHub: NVIDIA/OpenShell (star count read Oct 29, 2026)](https://github.com/NVIDIA/OpenShell)
- [X: a renamed form field silently broke a Zap](https://x.com/promo17062/status/2114324926673129914)

---
Source: https://cofounderbase.com/blog/trending-ai-agents-tools-week-2026-10-29/?utm_source=md&utm_medium=ai · Cofounderbase, a directory of agencies that build AI internal tools and automations. Data comes from each agency's own website and official partner directories; rankings are not sold.