# Robot Outlook agent guide

Purpose: compare an automation brief with source-backed installation reports and prepare supplier questions. Current coverage: food and drink palletizing, three vendor/integrator reports. This is not an independent performance test, a complete market search or a suitability certification.

## Web and JSON

Resolve these paths against the origin from which you read this guide:

- GET /search?q=URL_ENCODED_BRIEF: human-readable comparison.
- GET /api/research?q=URL_ENCODED_BRIEF: the same comparison as JSON.
- POST /mcp: read-only JSON-RPC MCP endpoint (HTTP JSON responses; no streaming session required).

The structured tool for the current published research is `investigate_automation`.

Example arguments:

```json
{"q":"Palletize food cases with slip sheets", "edited":"1", "rate":"12", "weight":"40", "height":"81"}
```

The same fields work as query parameters on /api/research and /search. `rate` is cases/minute; `weight` is pounds per case; `height` is inches. `edited=1` uses exactly the supplied numeric fields. Empty or omitted fields stay unknown. Without edited=1 the bounded parser extracts explicit quantities from the brief. Always review the parsed fields with the user before relying on the comparison. It does not understand every condition in prose. The query limit is 2,000 characters. Searches use URLs, so avoid sensitive plant or customer details.

## How to interpret results

- Check coverage before drawing conclusions. Unsupported tasks return no cases.
- Cite original source URLs and exact supporting passages. The browser source notes at /sources/{claim_id} provide preserved passages and capture dates.
- Vendor-reported facts are not independently verified.
- Keep required rates, claimed capacity, achieved production rate and robot trips separate.
- A matched dimension is a source-supported condition, not proof of overall suitability.
- A different condition signals an evidence gap; it does not prove impossibility.
- Preserve missing or conflicting evidence. Never fill it with guessed numbers.
- Washdown, freezer use, site layout and other unparsed constraints need explicit supplier confirmation.
- Return a concise assessment, relevant cases, differences, unknowns, and supplier questions. Do not produce an opaque robot ranking.

The response includes the versioned comparison method, normalized inputs, source snapshots and claim IDs. Save the response if you need a record of this assessment. Read-only searches do not publish or save a private brief into the corpus. The other MCP tools expose broader historical evidence and explicit-input economics; use investigate_automation for parity with the current website.


## AI employee deployment data

Robot Outlook also publishes a machine-readable deployment ledger for AI employees used around physical-world service work:

- GET /static/data/ai-employee-deployments.json: structured named deployments, tasks, outcomes, evidence tier and source URLs.
- GET /ai-employees/evidence/plumbing: computed human-readable synthesis for the current plumbing/home-services corpus.

Treat vendor-published customer case studies as deployment evidence, not independent performance validation. Prefer computed cross-case findings only when metric definitions and scopes are genuinely comparable.