Learn › Agent Harness, the big picture › Stop 7 › What a harness must do
Stop 7 of 9 · Pages
Receive the request
Before any thinking happens: where a request arrives, who sent it, and whether it should go any further.
- 1 · A front door where requests arrive
- 2 · Know who is asking
- 3 · Know what they are allowed to do
- 4 · Check what comes in
- 5 · A fair share for everyone
A front door where requests arrive
Also called: API server, gateway, channel adapters
Where does a message arrive, whatever channel it came from?
Why
Every channel speaks differently. One entry point turns them all into the same request, so everything behind it is written once.
Without it
Every channel grows its own copy of the agent, and they drift apart.
Think about
- Which channels come first, and which can wait
- The one shape of a request: who, what, and which conversation it belongs to
- How the reply travels back, and whether it appears word by word
In the big picture: The front door, with channel adapters in the Act family
Know who is asking
Also called: Authentication
Is this person really who they say they are?
Why
A model believes what it reads. So who the person is must be settled by ordinary sign-in before the model sees anything.
Without it
Anyone can type "I am a manager" and be treated as one.
Think about
- Customers, staff and other systems each sign in differently
- The identity must travel with the request, all the way to every tool
- What an anonymous visitor is allowed to do, if anything
In the big picture: The front door, and Identity and access in the Control family
Know what they are allowed to do
Also called: Authorization, permissions
What may this person see and change?
Why
An agent with one powerful account of its own can show anything to anyone. Tools should run with the authority of the person asking.
Without it
A polite request is enough to read another customer's data.
Think about
- Rights by role: customer, handler, manager
- Rights by record: your own policy, your own branch
- Which tools each role may use at all
In the big picture: Identity and access, in the Control family
Check what comes in
Also called: Input guardrails
Is this request safe to hand to the model?
Why
Some trouble can be caught before a single model call is paid for: an attempt to override the instructions, an identity number that should be masked, a request that is not this agent's job.
Without it
The model is the first and only line of defence.
Think about
- Which checks are simple rules and which need a small model
- What to do on a hit: block, mask, or let through with a flag
- How to say no to the person without teaching an attacker
In the big picture: The front door, and Guardrails in the Control family
A fair share for everyone
Also called: Rate limits, quotas, priorities, a queue
What happens when everyone asks at once?
Why
One busy user, or one script in a loop, can use up the model budget and the capacity for everybody else.
Without it
One runaway caller makes the service slow or unavailable for all.
Think about
- Limits for each person, each team and each agent
- Which work jumps the queue
- What a person is told when they are asked to wait
In the big picture: The front door, and Cost control in the See family