Before Notion agents work around the clock: build the control panel

The risky thing about AI agents is not that they sound futuristic. The risky thing is that they quickly start to feel normal.

An agent that summarizes customer questions, updates a Notion page, or watches a Slack channel does not feel like a major IT rollout. It feels like a helpful colleague. That is exactly why even a small team needs a simple front desk: who owns the agent, what may it read, what may it do, what does it cost, and when should it be stopped?

That is today's signal from Notion. During May, Notion has shipped a cluster of Custom Agent updates: a dedicated agent directory, new admin controls, spend visibility, Slack access for private channels, and a Developer Platform where developers and coding agents can sync data, build tools, and run code on Notion infrastructure.

Source: Notion 3.5: Notion Developer Platform

Source: New Custom Agent Directory

Source: New Custom Agent controls for admins

Source: Add Custom Agents to private Slack channels

This is not only interesting for tech companies. If you run a small business, association, consultancy, school, or team where work already lives in Notion, Google Drive, and Slack, the practical question is simple: how do we let AI help without filling the workplace with invisible little robots?

What Notion actually changed

Notion describes its new Developer Platform as a way for teams to connect any data, build agent tools, and bring in external agents without running their own servers. In practice, there are three useful pieces.

First, data from API-backed sources can be synced into Notion databases. An API is a structured way for software systems to talk to each other. For a smaller team, that could mean customer tickets, CRM records, subscription events, or internal lists landing in one place where people and agents can work on them.

Second, Notion introduced Workers. A Worker is Notion's hosted runtime for small pieces of specialized code. Notion points out that these tools can be more predictable than pure LLM reasoning and use fewer tokens when an agent needs to do the same thing repeatedly. That matters. Sometimes AI should write, prioritize, or reason. Sometimes it should just follow a rule.

Third, Notion is opening the door for external agents. Its examples mention Claude, Codex, and Decagon. That does not mean every small team should build an agent fleet tomorrow. It means the workspace is becoming a place where several AI helpers can show up. When that happens, order matters more than inspiration.

Source: Notion 3.5: Notion Developer Platform

The agent directory matters more than it looks

The most down-to-earth update may be the simplest one: Custom Agents now have their own place in Notion's Library. Teams can browse all workspace agents, pin favorites, and create new agents to automate recurring busywork.

That sounds like a product detail. For small organizations, it is a governance detail.

When an agent is hidden inside someone's personal workflow, it is hard to understand. When all agents are visible in a directory, the team can ask better questions:

• Which agents already exist?
• Who created them?
• Which ones work with customer data?
• Which ones can write back into Notion or Slack?
• Which ones run on a schedule?
• Which ones have not been used for a month?

This is the same shift we saw when shared documents replaced local files. The problem was not that people created documents. The problem was that nobody knew which version mattered. With agents, the version is not only a document. It can be an action.

Source: New Custom Agent Directory

Admin controls make agents less dramatic

Notion's new controls cover who can create agents, how much each agent may spend, how usage is tracked, and how an agent can be paused if it starts running away.

That may sound like enterprise language. A small business needs the same principle, only simpler. If two people can create agents, three people can invite them into Slack, and nobody checks cost until the invoice arrives, you have not introduced AI. You have introduced invisible operations.

Notion says admins can choose which individuals or groups may create Custom Agents. They can set per-agent credit limits with agent creators. At Enterprise level, they can also set a workspace-wide credit limit. The usage dashboard shows total usage broken down by agent, plus status and recent activity. If an agent is spending more than expected, it can be disabled while the creator adjusts it.

Source: New Custom Agent controls for admins

For a small school, the same idea translates like this: an agent that helps with weekly parent updates should not read student welfare notes. An agent that summarizes meetings should not write decisions. An agent that replies in an internal Slack channel should not be invited into private HR channels. This does not need to be complicated. It does need to be written down.

Slack access needs extra care

On May 1, Notion wrote that Custom Agents can read and reply in private Slack channels when the workspace enables access to private content and the agent is invited into the channel. Notion also says agents only see the private channels they are invited to.

Invitation-based access is good. But private Slack is often where small teams talk like humans: unfinished pricing, sensitive customer cases, people issues, frustration, negotiations, students who need support. An AI agent in that kind of channel is not just a tool. It becomes a new participant with memory, instructions, and possible actions.

So private channels should start as red zones. Not because AI can never be used there, but because someone must be able to explain why this specific agent needs this specific access.

Source: Add Custom Agents to private Slack channels

A simple agent register goes a long way

You do not need to start with a twenty-page policy. Start with an agent register in Notion, Airtable, Google Sheets, or whatever tool your team already uses.

Create one entry per agent. Use these fields:

• Agent name: what it is called in the workspace.
• Job: one sentence describing what it should do.
• Owner: one person, not "the team".
• Trigger: manual start, schedule, new Slack thread, new database row, or webhook.
• May read: exact pages, databases, channels, or files.
• May never read: for example HR, finance, student welfare, contracts, or private customer notes.
• May write: where the agent may create, update, or reply.
• Human review: what a person must approve before anything moves on.
• Cost limit: a simple weekly limit or credit limit.
• Stop rule: when the agent gets paused immediately.
• Last reviewed: date and a short note.

This is boring in exactly the right way. It makes agent work visible before it becomes expensive, embarrassing, or messy.

Copy the prompt: build your first agent control

Paste this into ChatGPT, Claude, Gemini, or Notion AI. Do not use real personal data or sensitive customer cases in the first test. Replace the brackets.

You are our AI coordinator. Help us build a simple agent register before we allow AI agents to work in Notion, Slack, or other internal tools.

Our organization: [short description]
Our team: [number of people and roles]
Tools we use: [Notion, Slack, Google Drive, CRM, email, other]
Three recurring tasks we want to automate:
1. [task]
2. [task]
3. [task]

Do this:
1. Suggest which tasks are suitable for a first AI agent and which ones we should wait with.
2. Create an agent card for the best first candidate with these fields: name, job, owner, trigger, may read, may never read, may write, human review, cost limit, stop rule, last reviewed.
3. List five risks that are realistic for a small team, not generic AI risks.
4. Suggest a one-week test where the agent may help, but may not publish, send, or change anything without approval.
5. Write a short instruction the owner can read every Friday to decide whether the agent should continue, change, or pause.

Keep it practical. If information is missing, state your assumptions clearly instead of making things up.

This exercise is useful even if you do not use Notion. The point is that the agent must not become a mystery. It should have a job card, an owner, and a stop button.

A first week you can actually run

If you want to test without turning it into a project, do this:

Day 1: Choose a harmless task. Good candidates are internal meeting summaries, suggested FAQ replies, knowledge base cleanup, or a draft weekly update. Bad first candidates are invoicing, HR issues, grading, legal contracts, or customer promises.

Day 2: Create the agent card. Write too narrowly rather than too broadly. "May read the page Support questions May" is better than "may read support".

Day 3: Run the agent on old material. Compare the result with how a human would have done it. Do not only look for grammar mistakes. Look for the wrong priority, wrong tone, missing caveats, and things the agent should have escalated.

Day 4: Let the agent help in a live workflow, but do not let it publish, send, or change anything without approval.

Day 5: Review three things: did it save time, did it create new controls, and is the owner willing to stand behind the result? If the answer to the last question is no, pause the agent or make it narrower.

This sounds slow. Good. Small teams do not have time to clean up after a helper that moves fast in the wrong direction.

What Hammer would build around this

For Hammer Automation, this is a classic Tool Forge problem. The tool is not only Notion, Slack, or an AI model. The tool is the whole workflow: data in, decision, human review, cost, action, and log.

In a practical rollout, we would start by mapping where the work happens today. Then we would pick one small agent with clear value and low risk. After that, we would build the register, ownership, and review around it. If the team first needs confidence in the working method, that belongs in Skill Forge: short exercises, better instructions, and shared rules for when AI may help.

The best AI agent in a small team is rarely the most impressive one. It is the one that does a clear job, in the right channel, with the right access, and that someone is willing to pause.

Common questions about Notion agents for small teams

Do we need Notion Developer Platform to start? No. Start with the agent register and a manual test week. Developer Platform becomes relevant when you want to sync your own data sources, build custom tools, or connect events from other systems.

Is the agent directory only for large companies? No. Small teams often need more visibility, not less, because the same person may handle sales, support, and admin. A directory makes it easier to see which AI helpers exist.

Should we let an agent reply in private Slack channels? Start with no. If you later open a private channel, write down exactly why the agent needs access, who owns it, and which information it must not use.

What is the smallest useful control? An owner, a read list, a write list, a cost limit, and a stop rule. Without those five things, the agent is too loose.