# AI agent infrastructure worksheet

A design worksheet from ipvolt. Public-documentation synthesis reviewed 17 September 2026; no provider benchmark or paid rental was run.

## Worked example: release-note monitoring assistant

Goal: read an allowlisted set of permitted release-note pages, summarize changes and save a source-linked report. The component choices below are a hypothetical architecture, not a deployed integration.

**Filled fictional record:** six permitted public release-note pages checked daily; a team runtime, scheduler, HTTP retriever and report store already work. Direct retrieval returns accepted text and the application detects changes. The missing component is text summarization. The model interface needs text messages and a text response; it does not require native tools, images, embeddings or `response_format`. Failed runs remain visible for follow-up. Model quality, context, data handling and availability still need evaluation.

**Derived choice:** retain the working runtime/retrieval/storage and evaluate a text-inference endpoint. This record establishes no need for a proxy purchase. The documented Proxies.sx text interface is a candidate for evaluation, not a tested recommendation. If required native `tools` or `response_format` are added to the record, its documented interface fails that new requirement; choose a compatible endpoint or evaluate a deliberate redesign.

| Requirement | Component you operate or buy | Evidence before you accept it | If missing |
| --- | --- | --- | --- |
| Run every agreed interval | Application scheduler/runtime | Run ID, start time, deadline and recorded completion/failure | Confirm where the scheduler runs; inference access alone is insufficient |
| Reach the selected source | Retrieval client; proxy only where needed | Intended URL/context and accepted response content | Diagnose retrieval before sending text to inference |
| Run a browser or arbitrary code | Browser/process/container or OS service | The service explicitly exposes that runtime | Do not substitute managed inference for OS access |
| Generate from accepted text | Model endpoint and inference adapter | Accepted request fields, assigned model, complete response and source support | Compare the actual model contract with your framework |
| Use model-native tool calls | Endpoint supporting the framework contract plus tool executor | End-to-end tool-call/result flow using the required request fields | An MCP connection elsewhere is not proof of model-native tool support |
| Require structured output | Compatible constrained-output interface or deliberately designed validation/rejection path | Valid and invalid output cases handled without accepting malformed results | Do not remove required schema fields and assume equivalent behavior |
| Save the result | Application storage/output validator | Report ID with source references and validation outcome | Text returned by the model is not proof of persistence |

For this example, compare the accepted text with the previous accepted snapshot. Record one final outcome: **source unchanged**, **report saved**, or **run failed at a named stage**. Keep retrieval, inference and storage failures distinct. These are design states; no monitor was executed for this worksheet.

## Fill this in before choosing services

- Job and permitted sources:
- Existing components that already work:
- Browser/runtime location and owner:
- Whether a proxy is required for this job, and why:
- Required model request fields:
- Model context/output requirements:
- Current availability and actual quote check:
- Data sent to each service and applicable retention/access terms:
- Acceptance condition and maximum end-to-end deadline:
- Separate retrieval/model/output failure records:

## The Proxies.sx example

Use the [proxy client portal](https://client.proxies.sx/) for its proxy-service starting point, the [compute overview](https://www.proxies.sx/compute) for managed inference scope and the [compute technical reference](https://agents.proxies.sx/compute/) for renter and supplier/runtime documentation. The current [compute API guide](https://www.proxies.sx/compute/api) says tools and response_format are unsupported. Verify the current contract before a purchase; these links do not establish tested performance.

Conceptual sources: [HTTP intermediaries](https://www.rfc-editor.org/rfc/rfc9110.html#section-3.7) and [MCP architecture](https://modelcontextprotocol.io/docs/learn/architecture).