Learn › Agent Harness, the big picture › Stop 7 › What a harness must do
Stop 7 of 9 · Pages
Act: tools, checks and people
The moment an agent stops talking and starts doing. This is where most of the guardrails live.
- 10 · Use tools
- 11 · Treat what tools return as data
- 12 · Check before acting
- 13 · Ask a person
- 14 · Check what goes out
Use tools
Also called: Tool hub, tool calling, MCP
How does the agent do anything in the real systems?
Why
A tool is an ordinary function with a name, a description and typed input. A shared hub means a tool is built once and every agent can use it.
Without it
The agent can talk about the claim but cannot register it.
Think about
- Tools that only read, and tools that change something
- A clear description, because the model chooses tools by reading it
- A common standard for sharing tools between agents and teams
In the big picture: Tool hub, in the Act family
Treat what tools return as data
Also called: Tool-result guardrails, fencing
What if a document the agent reads contains an instruction?
Why
Documents, emails, web pages and notes in records are written by other people. Anything in them can read like an instruction.
Without it
Anyone who can put text in front of the agent can steer it.
Think about
- Mark every tool result as data, every time
- Flag text that looks like an order
- Assume the fence will sometimes fail, and check actions as well
In the big picture: Guardrails, in the Control family
Check before acting
Also called: Action guardrails, pre-tool checks, policy
Should this particular action be allowed to run?
Why
This is the last moment before something real happens. Rules here are written in code, so a persuasive message cannot talk its way past them.
Without it
The business rules exist only as sentences in a prompt.
Think about
- Limits on amounts and on how many records one action may touch
- Whether the person actually asked for this
- What happens on a refusal: tell the model, ask a person, or stop
In the big picture: Guardrails, in the Control family
Ask a person
Also called: Human approval, human in the loop
Which actions need a person to say yes?
Why
Some actions are too large or too final for any agent to take alone. The agent prepares the work and a person makes the decision.
Without it
Either the agent does everything alone, or people check everything and nothing is saved.
Think about
- Which actions, at which autonomy level
- What the person is shown, so that the decision takes seconds
- The agent may wait hours for an answer, so its state must survive the wait
In the big picture: Human desk, in the Control family
Check what goes out
Also called: Output guardrails
Is this reply safe and true enough to send?
Why
A company answers for what its assistant says. A reply about cover, price or a refund should be backed by a record, not by the model's general knowledge.
Without it
A confident wrong answer becomes the company's word.
Think about
- Is every claim in the reply supported by something a tool returned?
- Does it show data about anyone else?
- Tone and scope: no promises it cannot keep
In the big picture: Guardrails, in the Control family