> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kvant.sbs/jeston/llms.txt
> Use this file to discover all available pages before exploring further.

# AI agents and model integrations

# AI agents and model integrations

***

## Platform · AI agents

> **Platform concern** — Build agent execution as a controlled workflow with tools, budgets, traces, and evaluation—not as an unbounded chat loop.

Model adapters, permissioned tools, durable state, streaming, and measurable quality.

### Key concepts

<Columns cols={3}>
  <Card title="Model" icon="sparkles">
    Probabilistic
  </Card>

  <Card title="Tool" icon="shield-check">
    Deterministic effect
  </Card>

  <Card title="Evaluation" icon="gauge-high">
    Quality signal
  </Card>
</Columns>

### Operating model

```mermaid placement="top-right" theme={null}
flowchart LR
    goal[User goal] --> plan[Model plan]
    plan --> policy[Tool policy]
    policy --> tool[Tool execution]
    tool --> evidence[Result and trace]
    evidence --> plan
    plan --> done[Bounded completion]
```

### Decisions to make before production

| Decision   | Recommended posture  | Failure it prevents                       |
| ---------- | -------------------- | ----------------------------------------- |
| Model      | Probabilistic        | Treat output as untrusted input.          |
| Tool       | Deterministic effect | Validate and authorize every call.        |
| Evaluation | Quality signal       | Keep a regression set and trace failures. |

<Note>
  Never let model text directly become a database query, shell command, permission decision, or external message.
</Note>

### Boundary and failure behavior

Agent workloads combine nondeterministic model output with deterministic application side effects.

### Questions for review

| Question                | Answer to document                                                                            |
| ----------------------- | --------------------------------------------------------------------------------------------- |
| What belongs in memory? | Only information with a retention, tenant, privacy, and deletion policy.                      |
| What should be traced?  | Prompt version, model, tool calls, latency, token usage, outcomes, and safe error categories. |
| When use a queue?       | When agent work must survive disconnects or may exceed request deadlines.                     |

### Implementation notes

| Engineering move          | Guidance                                                         |
| ------------------------- | ---------------------------------------------------------------- |
| **Define the capability** | State what the agent may decide and what it may never do.        |
| **Permission every tool** | Validate arguments, identity, tenant scope, and side effects.    |
| **Bound the loop**        | Set token, time, tool-call, and spend budgets.                   |
| **Evaluate behavior**     | Test groundedness, tool correctness, refusal, latency, and cost. |

### Decision lens

| Mode            | Practical emphasis                                            |
| --------------- | ------------------------------------------------------------- |
| **Interactive** | Stream plan and progress while keeping side effects explicit. |
| **Background**  | Persist intent and status so work survives disconnects.       |
| **Training**    | Version datasets, prompts, evaluators, and model settings.    |

## Related topics

<Columns cols={3}>
  * [stream · **Stream agent work**](/jeston/jeston/core/execution-and-streaming) — Read the focused guide for this boundary.
  * [shield-halved · **Protect tools**](/jeston/jeston/platform/security) — Read the focused guide for this boundary.
  * [gauge-high · **Benchmark agents**](/jeston/jeston/operations/benchmark) — Read the focused guide for this boundary.
</Columns>

## References

[1]: https://github.com/jeffersoncampos12p-dev/jeston "Jeston source repository"

[2]: https://www.npmjs.com/package/@kvantjs/jeston "Jeston package on npm"

[3]: https://nodejs.org/api/http.html "Node.js HTTP API"

[4]: https://developer.mozilla.org/en-US/docs/Web/API/AbortSignal "AbortSignal Web API"

[5]: https://react.dev/reference/react-dom/server "React server rendering APIs"

[6]: https://www.typescriptlang.org/docs/handbook/intro.html "TypeScript handbook"