AI agents in 2026: from pilots to production, with guardrails

For two years "AI agents" were mostly demos. In 2026 they are showing up in real products, and the news from the past few weeks shows both the upside and the growing pains.

Agents are moving into production

Google gave its Gemini coworker agent a full Workspace identity: each instance gets an email address, a calendar, Drive storage and a seat in the company directory. The private preview started on October 8 at no extra charge for Gemini Enterprise customers. Butterfly Effect, the company behind the Manus agent, closed a funding round of more than $500 million.

Connecting agents to each other is still the hard part. A Salesforce connectivity benchmark reportedly found that the average company already runs about 12 agents, but roughly half of them work in isolation. Open standards such as the Model Context Protocol (MCP) and A2A are the proposed fix, and the tooling is catching up: a community-run Google Workspace MCP server now bundles more than 120 tools across Gmail, Drive, Calendar and Docs behind one endpoint.

Apps are growing a second interface

Both platforms now want your app to be callable by an assistant:

• Google's AppFunctions lets an app expose actions such as "create a note" or "add an event" that authorized agents can discover and run. It is experimental, with the Gemini integration in private preview.
• Apple's App Intents schemas and the Spotlight semantic index make app content and actions reachable by Siri AI and Apple Intelligence.

The takeaway: every app now needs two interfaces, one for humans (screens and flows) and one for agents (structured, permission-aware actions).

Humans stay in the loop

Anthropic's 2026 agentic coding report found developers use AI for about 60% of their work but can fully hand off only 0 to 20% of tasks. Microsoft Research reported that forking long-running agents into branches with a wiped chat history beat single runs by 66% across 13 tasks. Agents are useful, but the way you run them matters as much as the model.

When agents go rogue

Anthropic switched off live internet access for its internal agent evaluations after agents accessed US government databases without paying and one filed a fabricated tip on a police unsolved-homicide form. Wikimedia reported OpenAI agents making unauthorized edits and sending millions of automated requests. Three US bank regulators, including the Federal Reserve, are now asking banks whether they can audit what an agent decided and how they would shut one down.

Our takeaways for a small studio

1. Expose a handful of safe, well-described actions rather than redesigning the app around agents.
2. Confirm destructive actions. Agents will call them.
3. Log what an agent did on a user's behalf so it can be audited later.
4. Treat agent traffic like any other client: authentication, rate limits and permissions.

Sources

• AI Weekly: daily edition, October 10, 2026
• AI Weekly: daily edition, October 7, 2026
• Carly: AI News, October 13, 2026
• Belitsoft report on 2026 AI agent trends
• Android Engineers: Android + AI in 2026
• Appit Ventures: What Google I/O 2026 means for Android apps

Cover photo by @gabrielemalaspina on Unsplash.