Learn › Agent Harness, the big picture › Stop 7
Stop 7 of 9 · Pages
What a harness must do
Before the big picture: everything a harness has to do, one need at a time. 33 requirements in seven groups, one small picture each.
The top row follows a request: it is received, the agent thinks, and the agent acts. These are the functional requirements, the things an agent does. The bottom row applies to every request at once: keep going, see, prove and change, and protect. These are the non-functional requirements, how well and how safely it does them.
Receive the request →Think →Act →Keep going →See →Prove and change →Protect →
All 33 requirements
| Requirement | Also called | The question an engineer asks | Group | |
|---|---|---|---|---|
| 1 | A front door where requests arrive | API server, gateway, channel adapters | Where does a message arrive, whatever channel it came from? | Receive the request |
| 2 | Know who is asking | Authentication | Is this person really who they say they are? | Receive the request |
| 3 | Know what they are allowed to do | Authorization, permissions | What may this person see and change? | Receive the request |
| 4 | Check what comes in | Input guardrails | Is this request safe to hand to the model? | Receive the request |
| 5 | A fair share for everyone | Rate limits, quotas, priorities, a queue | What happens when everyone asks at once? | Receive the request |
| 6 | Run the loop | Orchestration, the agent loop | What drives the turns, and when does it stop? | Think |
| 7 | Call the model, reliably | Model gateway | How does every agent reach a model, and what if the model is slow or down? | Think |
| 8 | Give the model the right context | Context building, retrieval, prompts | What exactly does the model get to see on each turn? | Think |
| 9 | Remember | Memory and state | What does it remember within a conversation, and across them? | Think |
| 10 | Use tools | Tool hub, tool calling, MCP | How does the agent do anything in the real systems? | Act |
| 11 | Treat what tools return as data | Tool-result guardrails, fencing | What if a document the agent reads contains an instruction? | Act |
| 12 | Check before acting | Action guardrails, pre-tool checks, policy | Should this particular action be allowed to run? | Act |
| 13 | Ask a person | Human approval, human in the loop | Which actions need a person to say yes? | Act |
| 14 | Check what goes out | Output guardrails | Is this reply safe and true enough to send? | Act |
| 15 | Survive a crash | Durable execution, checkpoints | What happens to a half-finished job when a server restarts? | Keep going |
| 16 | Do each action only once | Idempotency | What if the same step is tried twice? | Keep going |
| 17 | Cope when something fails | Time-outs, retries, fallbacks, circuit breakers | What does the agent do when a tool or the model is down? | Keep going |
| 18 | Be fast enough | Latency, streaming, caching | How long will a person wait, and what do they see while waiting? | Keep going |
| 19 | Serve many people at once | Scale, concurrency | Does it still work with a thousand conversations at the same time? | Keep going |
| 20 | Logs | Logging | What happened, and when? | See |
| 21 | Traces | Tracing, spans | What was the whole journey of this one request? | See |
| 22 | Numbers and alerts | Metrics, dashboards, alerting | Is it healthy right now, and who is told when it is not? | See |
| 23 | Know and cap the cost | Cost metering, budgets | What does one conversation cost, and what stops it running away? | See |
| 24 | An audit trail | Audit log | Who did what, why, and with whose approval? | See |
| 25 | Evals | Evaluations, golden sets, graders | Is it any good, and how would we know? | Prove and change |
| 26 | Attack it before others do | Red teaming, adversarial testing | Does the harness hold when someone tries to break it? | Prove and change |
| 27 | Release safely | Versioning, staged rollout, rollback | How does a change reach users without hurting them? | Prove and change |
| 28 | Change settings without changing code | Configuration | Can we change a limit without a new release? | Prove and change |
| 29 | Learn from every miss | Feedback loop | How does a mistake in production make the agent better? | Prove and change |
| 30 | Keep secrets away from the model | Secrets management | Where do passwords and keys live? | Protect |
| 31 | Protect personal data | Masking, data minimisation, retention | How much personal data does the model really need? | Protect |
| 32 | Keep customers apart | Isolation, multi-tenancy | Can one customer's data, memory or limits leak into another's? | Protect |
| 33 | Name an owner and a risk level | Governance | Who answers for this agent, and how much freedom does it have? | Protect |
Guardrails come in four kinds
"We need guardrails" is four different requirements, because there are four different places to stand.
| Where it stands | Name | What it catches | Group |
|---|---|---|---|
| On the way in | Input guardrails | Hidden orders, personal data, off-topic requests | Receive the request |
| On what the agent reads | Tool-result guardrails | Instructions hidden in documents, emails and records | Act |
| Before an action | Action guardrails | Payments, deletions and anything outside the person's rights | Act |
| On the way out | Output guardrails | Wrong claims, other people's data, promises the company cannot keep | Act |
The usual names for the non-functional requirements
If your team uses the standard words, this is where each one is covered.
Next: stop 8 puts all of these into one picture. Every box in it will be a requirement you have already met.