Proactive AI agents: what they do well, and the five things they still need

A proactive AI agent acts without waiting for a prompt. It watches your inbox, calendar or tools, notices something and takes the first step. That makes it useful for work, and it also means it acts with your accounts while you are not looking. Once you rely on one, you need five more things: approvals, a record of every run, corrections that stick, repeatable workflows, and one place to run every agent.

What is a proactive AI agent?

A reactive assistant waits. You type, it answers, and between prompts it does nothing. A proactive agent watches for a trigger, such as a new email, a calendar change or a failing check, and decides whether to act on it.

Consumer AI moved this way in the last year. OpenAI's ChatGPT Pulse, launched as a preview in September 2025, researches overnight from your chats and, if you connect them, your Gmail and Google Calendar, and has a short brief ready each morning. When Meta launched its Muse personal agent in September 2026, the Hacker News thread reached 666 points and 742 comments. The pitch is the same everywhere: an assistant that gets to things before you do.

What are proactive agents good at?

They are good at the first move on work that arrives on its own schedule:

• Morning briefs. Meetings with context, replies owed, what changed overnight.
• Inbox triage. Sorting, labelling and drafting replies for the messages that need you.
• Follow-ups. Noticing a lead that went quiet or a promise that is due.
• Monitoring. Watching a metric, a site or a repository and flagging the change that matters.
• Research. Pulling together what is new on a topic before you ask.

For one person's day, that is often enough. The trouble starts when the agent's work has to hold up for a team, a client or a customer.

Why isn't a proactive agent enough for work?

A commenter on the ChatGPT Pulse launch thread (627 points, 730 comments) put the split precisely: "I generally don't want my AI to be proactive at a personal level despite it being useful at a business level."

At the business level, an agent that acts on its own raises questions a personal assistant never has to answer. What is it allowed to do without asking? What did it actually do yesterday? Does it stop making the mistake you corrected last week? Does it do the Monday report the same way every Monday? And when there are six of these agents, who is in charge of them?

Those five questions are the five things a proactive agent still needs.

1. Governance: what may it do without asking?

A proactive agent works through your accounts. Reading is low risk; sending, posting, deleting and paying are not. The working rule is simple: let the agent do reversible work inside its own space, and make anything that leaves that space, or cannot be undone, wait for a person.

The risk is real. A safety study accepted at ICML 2026 examined the most widely deployed personal AI agent of early 2026, which "operates with full local system access" and connects to Gmail, Stripe and the filesystem. Poisoning a single part of the agent's stored state raised the average attack success rate from 24.6% to between 64% and 74%. AI agent governance for small teams covers approvals and access in detail.

2. Observability: what did it actually do?

A proactive agent that worked overnight leaves you a summary. A summary is a claim. For work you hand to someone else, you need the record behind it: what triggered the run, which tools it used, what it produced, and whether that output met the bar you set.

Without that record, a run that quietly did nothing looks the same as a run that did the job, and you only find out when someone asks where the work went.

3. Refinement: does a correction stick?

Every agent gets things wrong at first. What matters is whether a correction carries into the next run. In a chat, it carries only as long as that conversation does. For recurring work, a correction has to land somewhere permanent: in the agent's instructions, in a checklist, or in a memory scoped to that job, so the same note never has to be given twice.

Self-improving AI agents covers how that feedback loop works and how to keep it under your control.

4. Workflows: does it do the same job the same way?

An agent decides its own steps each time. That flexibility is the point for an unfamiliar task, and a liability for a repeat one. A Monday report should come out in the same shape every Monday, from the same sources, to the same place.

That is what a workflow adds: the job written down once, with what starts it, what it produces and what counts as done. The agent's judgment still does the work inside each step. AI agents vs workflows covers when to use each and how to combine them.

5. One orchestrator: who runs the agents?

One proactive agent is easy to watch. Five are not. Each has its own chat, its own notifications and its own idea of what it is doing, and you become the person relaying between them. Gartner expects the average Fortune 500 company to go from fewer than 15 AI agents in 2025 to more than 150,000 by 2028.

The fix is one lead agent that plans the work, hands each piece to the right agent, checks what comes back, and brings you only the decisions. Too many AI agents covers the consolidation.

How do you cover all five?

Whatever tools you use, the setup looks like this:

1. Approvals for anything that leaves. Sending, posting, publishing, deleting and paying wait for a person. Everything else runs.
2. A record of every run. What started it, what it did, what it produced, and a link to the output.
3. Corrections written down. Every repeated correction goes into the agent's instructions or the job's checklist, where the next run will read it.
4. Repeat jobs saved as workflows. Each with its trigger, its output and a checklist of what done means.
5. One lead and one inbox. One agent that assigns the work, and one place where every question that needs you arrives.

How does Kavela cover all five?

Kavela is built around a lead agent, Kav, that runs every other agent in the workspace.

• Governance. You choose how much Kav does alone: Assistive checks with you before any change, Supervised handles its own workspace and checks before reaching outside it, and Autonomous works on its own while still checking the first time it uses a connection. Deleting files or workflows always needs your approval. Every question lands in one Approvals inbox.
• Observability. Every workflow run is checked by a second agent against your checklist, line by line, and the result shows on the run with what missed.
• Refinement. A comment on a result goes straight back to the agent. Your approvals and corrections are kept as memory scoped to that agent and job, and when Kav sees the same correction more than once it teaches the agent a lasting lesson.
• Workflows. A job written in plain words becomes a workflow that runs on a schedule, on an event from another service, or when you press Run, with every change saved as a new version.
• One orchestrator. You tell Kav the outcome. Kav plans it, assigns an agent to each part, has a new one built when none fits, and puts every task on one calendar.

Each agent works in its own cloud computer with a real browser, on your own Claude or Codex plan.

Kavela starts with a 14-day free trial, no card. Start your free trial