Learn › Agent Harness, the big picture › Stop 7 › What a harness must do
Stop 7 of 9 · Pages
Think: run the loop and call the model
The part people picture when they say "agent": the loop, the model, and what the model is given to work with.
Run the loop
Also called: Orchestration, the agent loop
What drives the turns, and when does it stop?
Why
Something has to send each request, read each response, run each tool and decide when enough is enough. The model does none of that.
Without it
The loop runs until someone notices the bill.
Think about
- Single call, workflow or agent: the simplest kind that does the job
- The limit on turns, on time and on spend
- One agent, or a router that hands work to several
In the big picture: Orchestration, the band above the agents
Call the model, reliably
Also called: Model gateway
How does every agent reach a model, and what if the model is slow or down?
Why
Model services slow down, fail and change. If each agent calls a model its own way, each one has to solve that again.
Without it
One outage at a model provider stops every agent at once.
Think about
- Which model for which job: a small fast one, or a strong one
- How long to wait, how often to retry, and what to fall back to
- Keys in one place, and tokens counted in one place
In the big picture: Model gateway, in the Think family
Give the model the right context
Also called: Context building, retrieval, prompts
What exactly does the model get to see on each turn?
Why
Stop 4 showed that our code sends everything on every turn. Choosing what goes in is the strongest control we have over what comes out.
Without it
The model answers from general knowledge, confidently and wrongly.
Think about
- What is always sent, and what is fetched only when needed
- The order, and a size budget, because every word costs money and attention
- What must be kept out, because this person may not see it
In the big picture: Knowledge, in the Think family
Remember
Also called: Memory and state
What does it remember within a conversation, and across them?
Why
Between turns and between days, the model holds nothing. Anything that should last has to be written down by our code and sent back in.
Without it
Every conversation starts from zero, or sends so much that it becomes slow and costly.
Think about
- Within one conversation, and across conversations: two different stores
- Who may read a memory, and for how long it is kept
- What to drop when the conversation grows too long to send
In the big picture: Memory and state, in the Think family