Home › Key concepts › The harness
Key concept · Walk (intermediate) · the same on every platform
The harness
Everything around the model that makes it safe to use.
What it is
The harness is the ordinary software around a model call: building the context, the tools, the checks before and after, state, retries, approvals, tracing, evaluation and cost limits. Most of an AI system is harness, not model.
Why it matters
The interesting failures are almost never in the model. They are in the harness, and the harness can be specified and tested like any software.
An example
Swap the model and the harness stays; the same tests still decide whether the system is safe to ship.
The common mistake
Spending months on prompts and none on the checks, state and tests around them.
Where you meet it on each platform
The idea stays the same; only the product name changes.
| Job | AWS | Azure | Google Cloud | Databricks | Open source |
|---|---|---|---|---|---|
| Agent runtimewhere a long-running agent keeps its state and waits safely | Bedrock AgentCore | Foundry Agent Service / Durable Functions | Vertex AI Agent Engine | Mosaic AI Agent Framework | DBOS or Temporal + your loop |
| Guardrailsblock what must never go in or come out | Bedrock Guardrails | Azure AI Content Safety | Model Armor | AI Gateway guardrails | rule checks in your own code |
| Tracingsee every model call, tool call, token and cent | CloudWatch + X-Ray (OpenTelemetry) | App Insights (OpenTelemetry) | Cloud Trace (OpenTelemetry) | MLflow tracing | Phoenix or Langfuse |
Related ideas
Where it shows up
Part of the Walk level on these paths: AI Engineering (tools and agents) · Forward Deployed Engineering (scale).