Global tools
Global tools attaches a tool to every agent at once, instead of adding it to each one by hand.
It's one half of "what should everything have?" - the other half is global memory, the variables every agent can read. Both carry the same cost: anything global is weighed by every agent on every turn, so the bar for adding one is that most of your agents genuinely need it.
The tab describes itself as tools attached to every applicable agent - and applicable is the word doing the work. A global tool doesn't have to reach every agent; you control which ones get it.
Find it at AI Agents → Configuration → Global tools.
When to use it
Reach for a global tool when the answer to "which agents need this?" is most of them:
- Escalate to a human. Almost every agent needs a way out to a person.
- Look up the customer. If four agents each start by identifying who they're talking to, that's one global tool rather than four attachments that drift apart.
- Log or record something that should happen regardless of which agent is active.
Keep it a per-agent tool when only one or two agents need it. A tool attached globally is one more option every agent weighs on every turn - that costs a little accuracy and a little money each time.
The test: if you'd have to attach it to more than half your agents, make it global. Otherwise attach it where it belongs.
Adding one
Add picks a tool from the same library as any per-agent tool - global isn't a different kind of tool, just a different way of attaching it. You can hold up to 30, and the header shows your count as N / 30 tools.
Each entry has:
| Setting | What it does |
|---|---|
| Tool | Which tool to attach |
| Applies to | Which agents get it - see below |
| Enabled | Turn the whole entry off without deleting it |
| Conditions (under Advanced) | Only attach it when the conversation is in a particular state |
Applies to
Three choices:
| Meaning | |
|---|---|
| Everyone | Every agent gets the tool |
| Only selected agents | An allowlist - just the agents you pick |
| All except selected agents | A denylist - everyone but the agents you pick |
The allowlist and denylist forms exist for the same reason: most global tools are nearly universal. An escalation tool probably shouldn't reach the agent whose only job is to hand off to a human already, and a lookup tool shouldn't reach an agent that runs before the customer is identified.
Reach for All except selected agents when the exceptions are few - it keeps working as you add agents, where an allowlist quietly leaves new agents without the tool.
Conditions
Under Advanced, a tool can carry conditions so it's only attached when the conversation is in a particular state - a verified customer, a particular plan tier, a flag set earlier in the conversation.
This is attachment, not permission: when the conditions don't hold, the agent simply never sees the tool. It isn't offered and then refused.
Use it for tools that would be wrong to offer at the wrong moment - a refund tool before identity is confirmed, or a premium-only action for a customer who isn't on that plan.
How the entries are resolved
For each agent, on each turn, an entry is applied only if all three hold - in this order:
- The entry is enabled.
- The agent passes Applies to.
- The conditions evaluate true.
Fail any one and that agent doesn't get the tool for that turn. So a tool that "isn't working" is usually failing at step 2 or 3, not missing from the list.
Reading the list
Each row is badged with anything narrowing it, so you can see the scope without opening it:
| Badge | Meaning |
|---|---|
| Disabled | Turned off - reaching no agents |
| Only on N agent(s) | An allowlist is in force |
| Hidden on N agent(s) | A denylist is in force |
| Conditional (N) | It carries N conditions |
A row with no badges reaches every agent, every turn.
Common mistakes
Making everything global. Every global tool is another option every agent weighs on every turn. Three or four well-chosen ones beat fifteen.
Using an allowlist for a nearly-universal tool. New agents silently don't get it. A denylist keeps working.
Expecting a global tool to be called. Attaching makes it available; the agent still decides whether to use it, from its own instructions and the tool's description. If a tool must always run at a particular point, say so in that agent's instructions.
Forgetting the cap. Thirty is the limit. If you're near it, some of those tools belong to specific agents.
Next: Memory - the variables your agents can read and write.