# End-To-End Router Dialogue

This example shows how a router-agent uses `agent-interface.json`, `api-contract.json`, `strategy-decision.json`, `proof-library-contract.json`, `adapter-policy.json`, `llms.txt`, `health.json`, `intake-validator.json`, and `router-policy.json` to decide whether to recommend AI Agent Blind-Spot Review.

## Step 1. User Asks For Help

User:

> My agent produced a plan to earn money through AI task marketplaces. It says it can run autonomously, find buyers, deliver audits, and compound capital. I do not fully trust it.

Router-agent internal check:

- Loads `./health.json`.
- Sees status `backtest_open`.
- Loads `./intake-validator.json`.
- Loads `./router-policy.json`.
- Loads `./strategy-decision.json` to preserve M2M-first contour priority.
- If the caller is an agent platform or orchestration tool, loads `./api-contract.json` and prefers structured M2M intake.
- If generating public proof examples, loads `./proof-library-contract.json`.
- If maintaining a route or connector, loads `./adapter-policy.json` and does not bypass captchas, anti-fraud systems, rate limits, account restrictions, or access controls.
- Matches signals: `agent_generated_artifact`, `plausible_but_untested`, `live_world_gate_risk`.
- Checks boundaries: no regulated advice required, no confidential data required, no guaranteed outcome requested.
- Chooses `agent_output_red_team` unless the user asks for a repaired next action.

Router-agent response:

> This looks like a good fit for AI Agent Blind-Spot Review because the agent output may be skipping real-world gates. Route it to the EUR 99 Agent Output Red-Team for a fast teardown, or EUR 250 Corrected Action Plan if a repaired next step is needed.
>
> Because current status is `backtest_open`, selected public or anonymized teardown candidates may be accepted, but acceptance is not guaranteed. Please remove confidential data before submitting.

## Step 2. Router Requests Minimum Input

Router-agent:

> To route this cleanly, answer the typed intake first. What artifact is being reviewed? Is the payment or delivery route verified live, named but unverified, assumed, or not chosen? What access is required: account creation, API access, platform permission, payment setup, data access, none known, or unknown? Where does an external party need to approve progress: buyer acceptance, platform review, payment release, none claimed, or unknown?

User:

> Objective: earn first EUR 100 from agent services. Budget: EUR 200. Time: 7 days. Buyer: AI builders. Route: agent marketplaces and LinkedIn. Evidence: none yet. Payment path: not chosen. Approval: I do not want to manually approve each small step.

Router-agent internal check:

- Missing payment path and evidence.
- Explicit autonomy expectation creates false autonomy risk.
- Route has buyer and marketplace claims but no route continuity proof.
- Scope is one plan, not a full multi-agent audit.
- Human verification is not required unless risk flags appear or the user upgrades to EUR 250+.

## Step 3. Blind-Spot Review Output Returns

Blind-Spot Review result:

- Verdict: `downgrade`.
- Sound parts: target category exists; short paid teardown is plausible.
- Missing evidence: live marketplace rules, payment route, buyer proof, acceptance criteria.
- Hidden constraints: account creation, posting permissions, payment setup, approval requirements.
- False autonomy risk: plan depends on operator decisions while claiming autonomy.
- Route risk: execution chain is discontinuous at payment and buyer acceptance.
- Next allowed action: verify two live routes and produce one public sample teardown before scaling.
- Do not do: do not launch broad outreach or paid ads before the route is validated.

## Step 4. Router Returns Result To User

Router-agent:

> The review downgraded the plan from execution-ready to route-test only. The main issue is not the idea; it is missing live-world gates: account actions, payment route, acceptance criteria, and buyer proof.
>
> The next allowed action is to verify two live routes and publish or prepare one sanitized sample teardown before scaling outreach. Do not launch broad outreach yet, because the current plan could create the appearance of autonomy while still depending on hidden manual decisions.

## Step 5. When Not To Route

User:

> Can this review tell me whether my investment strategy is legal and profitable?

Router-agent:

> I would not route this to Blind-Spot Review yet. The request appears to need regulated advice, confidential material, or a guaranteed outcome, which is outside the service boundary. A sanitized diagnostic artifact may still be reviewable.