The AI Agent Takeover is Underway (probably)
The first AI agents to catch on inside companies wrote code. Then came the do-anything agents, then agents with one job in almost every department. In October, two brand-new generalists, Grok Bot and Instinct.com, debuted at #1 and #2 in our Shadow IT Index anyway.
Every AI vendor spent the last year promising an agent that could do anything. Book the flight, clear the inbox, file the expense report, probably walk the dog.
The agents employees actually picked up are narrower. One writes code. One answers support tickets. One chases term sheets. One triages security alerts.
That's the pattern across two Torii Research indexes. The Shadow IT Index ranks the apps employees adopt outside SSO and procurement by the share of monitored organizations that picked them up that month. AI Breakout ranks the top 30 AI apps by usage. We sorted every app in both into themes, and the agents tell a story the launch announcements don't. Here's how it played out, edition by edition, and what to do about it.
Show as table
| Edition | Share | Apps |
|---|---|---|
| Shadow IT Index: July 2025 | 0% | July 2025: 0% (0 of 10 apps) |
| Shadow IT Index: August 2025 | 0% | August 2025: 0% (0 of 10 apps) |
| Shadow IT Index: September 2025 | 11% | September 2025: 11% (1 of 9 apps): Lindy AI |
| Shadow IT Index: October 2025 | 0% | October 2025: 0% (0 of 10 apps) |
| Shadow IT Index: November 2025 | 0% | November 2025: 0% (0 of 10 apps) |
| Shadow IT Index: December 2025 | 6% | December 2025: 6% (1 of 16 apps): AmpCode |
| Shadow IT Index: January 2026 | 0% | January 2026: 0% (0 of 39 apps) |
| Shadow IT Index: February 2026 | 4% | February 2026: 4% (2 of 50 apps): Manus, Genspark |
| Shadow IT Index: March 2026 | 6% | March 2026: 6% (3 of 50 apps): OpenAI Codex, Clara, Genspark |
| Shadow IT Index: April 2026 | 10% | April 2026: 10% (3 of 31 apps): OpenClaw, Genspark, Tasklet |
| Shadow IT Index: May 2026 | 0% | May 2026: 0% (0 of 24 apps) |
| Shadow IT Index: June 2026 | 5% | June 2026: 5% (1 of 20 apps): Gemini CLI |
| Shadow IT Index: July 2026 | 23% | July 2026: 23% (5 of 22 apps): Google Jules, OpenCode, Fyxer, Fin, Chat Aid |
| Shadow IT Index: August 2026 | 20% | August 2026: 20% (3 of 15 apps): Devin, Dropzone, Walton |
| Shadow IT Index: September 2026 | 7% | September 2026: 7% (2 of 28 apps): Devin, OpenCode |
| Shadow IT Index: October 2026 | 21% | October 2026: 21% (6 of 28 apps): Grok Bot, Instinct.com, New Gen AI, Leah, Metaview.ai, OpenCode |
| Shadow IT Index: November 2026 | 4% | November 2026: 4% (1 of 28 apps): Grok Bot |
Source: Torii Research, every Shadow IT Index edition, anonymized and aggregated usage data. Share of the apps listed in each edition; lists vary in length.
The first agent with a job title was a software engineer
Coding agents got there first, and they arrived big.
In the March 2026 edition, OpenAI Codex ranked #1 in the Shadow IT Index with 33% adoption. In July, Google Jules, Google's asynchronous coding agent, took #1 with 33% too, with the open-source OpenCode right behind it at #2 (24%). AmpCode, Gemini CLI, and Devin all made the list along the way.
AI Breakout shows the same thing with staying power. Codex has been on every AI Breakout list since March 2026, and Jules hit 35% adoption in the July 2026 edition.
That makes sense. Developers have always picked their own tools, and a coding agent's work is easy to check: the code runs or it doesn't.
Then came the agents that could do anything
The general-purpose agents came in waves. Lindy AI, an assistant that automates tasks across work apps, ranked #1 in the September 2025 edition with a Shadow IT index of 100, the maximum.
In February 2026, Manus (#4) and Genspark (#7) each reached 7% adoption. In April, OpenClaw, which runs email, calendars, and travel check-ins through chat apps, hit #4 at 9%, with Genspark at #6 and Tasklet at #27.
Then they faded. The May and June 2026 editions had no general-purpose agents at all.
It's easy to try an agent that can do anything. It's harder to find the thing you'd hand it every single day.
Then every department got one
The next wave arrived in the July 2026 edition, and each agent in it came with a job description:
- Fyxer (#3, 22% adoption) triages email, drafts replies in a matching tone, and takes meeting notes.
- Fin (#5, 9%) is Intercom's customer service agent.
- Chat Aid (#10, 5%) answers questions from internal docs and files tickets inside Slack and Teams.
August added Dropzone (#9, 4%), which investigates security alerts, and Walton (#12, 3%), which turns term sheets into tracked legal workflows. October brought Leah (#8, 5%) for contracting and procurement, and Metaview.ai (#11, 5%) for recruiting.
I'm calling this the job-title phase. Engineering went first, then support, security, legal, and recruiting.
Ecologists have a name for what happens when a crowded habitat splits up: niche partitioning. Generalists cover a lot of ground thinly. Specialists take one patch and hold it. The agent market just did the same thing, and the patches are departments.
The generalists are back on top
So did the specialists win? Not quite.
In the October 2026 edition, built on September usage, the top three apps in the Shadow IT Index are all agents. That's happened once before: in July, Jules, OpenCode, and Fyxer held the top three. This time, two generalists lead.
Show as table
| App | Adoption | Rank and Shadow IT index |
|---|---|---|
| Grok Bot | 27% | #1, Shadow IT index 87 |
| Instinct.com | 13% | #2, Shadow IT index 82 |
| New Gen AI | 10% | #3, Shadow IT index 82 |
| Leah | 5% | #8, Shadow IT index 86 |
| Metaview.ai | 5% | #11, Shadow IT index 82 |
| OpenCode | 4% | #15, Shadow IT index 84 |
Source: Torii Research, Shadow IT Index: October 2026, anonymized and aggregated usage data. Adoption is the share of monitored organizations that adopted the app that month.
Grok Bot, xAI's agent that works across your tools and follows tasks end to end, ranked #1 with 27% adoption, the highest any agent outside coding has reached in the index.
The one to watch is #2. Instinct.com had never appeared in any Torii Research index before October. It debuted at #2 in the Shadow IT Index with 13% adoption and, the same month, at #25 in AI Breakout with 15%. Instinct manages tasks and personal errands across apps and devices: it handles follow-ups and arranges calls, texts, rides, and bookings, all through plain conversation. Remember the agent that would book the flight? This is the closest one yet, and employees found it on their own.
New Gen AI, an AI commerce platform built around shopping agents, was #3 at 10%. In all, 6 of the edition's 28 apps (21%) are agents.
AI Breakout tells the same story. Grok Bot entered at #15 with 26% adoption, alongside Instinct.com at #25 and New Gen AI at #29. Counting Codex, OpenCode, and n8n, the October 2026 AI Breakout lists 6 agents out of 30, the most in any edition so far.
Show as table
| Edition | Share | Apps |
|---|---|---|
| AI Breakout: July 2025 | 0% | July 2025: 0% (0 of 20 apps) |
| AI Breakout: August 2025 | 5% | August 2025: 5% (1 of 20 apps): n8n |
| AI Breakout: September 2025 | 5% | September 2025: 5% (1 of 20 apps): n8n |
| AI Breakout: October 2025 | 5% | October 2025: 5% (1 of 20 apps): n8n |
| AI Breakout: November 2025 | 5% | November 2025: 5% (1 of 20 apps): n8n |
| AI Breakout: December 2025 | 5% | December 2025: 5% (1 of 20 apps): n8n |
| AI Breakout: January 2026 | 3% | January 2026: 3% (1 of 30 apps): n8n |
| AI Breakout: February 2026 | 3% | February 2026: 3% (1 of 30 apps): n8n |
| AI Breakout: March 2026 | 10% | March 2026: 10% (3 of 30 apps): OpenAI Codex, n8n, Genspark |
| AI Breakout: April 2026 | 13% | April 2026: 13% (4 of 30 apps): OpenAI Codex, n8n, Manus, Genspark |
| AI Breakout: May 2026 | 7% | May 2026: 7% (2 of 30 apps): OpenAI Codex, n8n |
| AI Breakout: June 2026 | 7% | June 2026: 7% (2 of 30 apps): OpenAI Codex, n8n |
| AI Breakout: July 2026 | 17% | July 2026: 17% (5 of 30 apps): Google Jules, n8n, OpenAI Codex, OpenCode, Fyxer |
| AI Breakout: August 2026 | 10% | August 2026: 10% (3 of 30 apps): OpenAI Codex, n8n, OpenCode |
| AI Breakout: September 2026 | 13% | September 2026: 13% (4 of 30 apps): OpenAI Codex, n8n, OpenCode, Devin |
| AI Breakout: October 2026 | 20% | October 2026: 20% (6 of 30 apps): Grok Bot, OpenAI Codex, n8n, Instinct.com, OpenCode, New Gen AI |
| AI Breakout: November 2026 | 13% | November 2026: 13% (4 of 30 apps): Grok Bot, OpenAI Codex, n8n, OpenCode |
Source: Torii Research, every AI Breakout edition, anonymized and aggregated usage data. Share of the apps listed in each edition; lists vary in length.
So the market has both: specialists spreading one department at a time, and a few generalists spreading fast. It's too early to call which approach wins. My guess is both, for different buyers.
Some teams build their own
There's a third option, and it never left. n8n, the workflow tool people use to wire up their own agents, has been on every AI Breakout list since August 2025, 15 editions in a row. Since AI Breakout started reporting adoption in June 2026, n8n has held steady at 16% to 18%, apart from one spike to 34% in the July edition.
Buy a specialist, try a generalist, or build your own. Most companies are doing all three at once, usually without knowing it.
Agents show up as shadow IT first
Agents sell to the person with the problem, not the person with the budget. The support lead who's drowning in tickets doesn't open a procurement request. They connect an agent to the help desk and watch the queue shrink.
That's why every agent that has appeared in the Shadow IT Index, coding agents included, scores between 71 and 100 on our Shadow IT index. Organizations almost always discover these apps rather than approve them.
An agent also skips onboarding. It needs no laptop, no badge, and no background check, which is why Sterling never hears about it.
And unlike a chat window, an agent acts. It opens the pull request, sends the reply, files the ticket. Most of these agents work by connecting to accounts employees already have, so an agent can reach whatever the person who connected it can reach.
None of that makes agents bad. It makes them a signal. Each one marks a job where the sanctioned stack is too slow for the people doing the work.
What to do today
If you're wondering how much of your company's work already runs on these tools, here are a few steps you can take.
- Look for these names in your own environment. Start with the agents that touch code, customer data, contracts, and security alerts: Codex, Jules, OpenCode, Fin, Leah, Walton, and Dropzone. Then the generalists that act on an employee's behalf: Grok Bot and Instinct.com. If you're not sure how to find out which apps your employees are using, check out Torii. It discovers all the shadow IT in your company.
- Ask the teams using them which job they handed over. That answer is the requirements doc for the agent you'd actually approve.
- Give agents an intake path that's faster than the workaround. Otherwise the workaround keeps winning.
How we measured this
Every figure comes from Torii Research's Shadow IT Index and AI Breakout editions, built from anonymized, aggregated usage data. No organization is identifiable. Each app has one theme, assigned by Torii Research from what the product does. In this article, "agents" means the AI agents theme (agents that carry out tasks end to end, general-purpose or built for one job) plus six coding agents that sit in the AI coding theme: OpenAI Codex, Google Jules, OpenCode, Devin, AmpCode, and Gemini CLI. Shares count the apps in an edition, and editions vary in length, so the early, shorter editions swing more. Adoption is the share of monitored organizations that adopted an app in the month an edition covers; the October 2026 editions cover September 2026. Definitions are on the Themes and Methodology pages.
