Learn › Agent Harness, the big picture › Stop 8
Stop 8 of 9 · Page
The architecture, simple then zoomed in
The whole system in five bands and two rails. Then one band opened up into five families of concerns.
First, the simple picture
Read it from the top as one request travelling down.
- People ask. Customers, partners and staff, on whatever channel they already use.
- The front door checks every request before any agent sees it: who is asking, is anything hidden in the message, is it in scope.
- Agents. One for each business job. Each has its own autonomy level from stop 2.
- Shared parts. Everything an agent needs that is not specific to its job. Built once, used by every agent.
- Business systems. The real records. Agents reach them only through tools, and only with the authority of the person asking.
Two rails run alongside. Lifecycle is the path every agent takes from idea to production and back round through what it learned. Governance is who owns the agent, how risky it is, and what the law requires.
Now zoom in on the shared parts
Twelve parts is too many to hold in your head, so they are grouped into five families. Each family answers one question that a business will ask about any agent.
| Family | The question it answers | Parts | Without it |
|---|---|---|---|
| Think | What does it know, how does it decide? | Model gateway, Knowledge, Memory & state | The agent guesses from what the model happens to remember |
| Act | What can it do? | Tool hub, Channel adapters | The agent can talk but cannot do anything |
| Control | Is it safe and allowed? | Identity & access, Guardrails, Human desk | The six stories in stop 5 |
| See | What did it do, what did it cost? | Observability, Cost control | Nobody can say what happened, or what it cost |
| Change | Is it good? Can we change it safely? | Eval library, Release & config | Every improvement is a gamble, so nothing gets improved |
Where each requirement lives
Stop 7 listed what a harness must do. Here is where each of those requirements sits in the picture.
Where the earlier stops land
| What you saw | The family that owns it |
|---|---|
| The request and response in the loop (stop 4) | Think: the model gateway sends it; memory and state hold the messages |
| The tool our code ran (stop 4) | Act: the tool hub |
| The check before the payout, and the handover to a person (stop 5) | Control: guardrails and the human desk |
| The tool running with the user's authority (stop 5) | Control: identity and access |
| The limit on turns and on spend (stops 4 and 5) | See: cost control |
| The record of the refused payout | See: observability |
| Knowing the fix still works after the next change | Change: the eval library and release |
Go deeper when you are ready
The same architecture exists at more zoom levels in the Agent Harness tutorial. Nothing new is added there; each picture opens one more level.
The full abstract picture: seven bands, two railsThe five families, part by partA north star, not a checklist
Next: stop 9 follows one real message through all of it.