Overview

The Open Hiring Harness is designed to be AI-native. Agents should be able to:

Read structured resources (profile, availability, services)
Request permissioned access (rates, detailed availability, reputation)
Invoke actions (request a quote, hold time, submit an endorsement)

All interactions respect the policies declared in the harness.

Resources (read-only)

Agents can read public data as MCP resources:

Resource URI	Description
`harness://identity`	Public identity (handle, display name, tagline, links)
`harness://offer/roles`	Roles and specialties
`harness://offer/services`	Service catalogue
`harness://availability/summary`	Availability status (open/limited/closed)
`harness://proof`	Case studies and credentials

Tools (actions)

Agents can invoke tools to perform actions:

Tool	Description	Requires consent
`request_access`	Request permissioned data (rates, detailed availability)	—
`request_quote`	Request a quote for a specific service	Yes
`hold_time`	Request a tentative booking hold	Yes
`submit_endorsement`	Submit an endorsement for review	Yes

Consent flow

When an agent requests permissioned data or invokes a protected tool:

Agent calls request_access with identity, purpose, and requested scopes
Harness owner receives the request (via configured notification channel)
Owner grants or denies; if granted, a consent receipt is issued
Agent can now access the scoped data until the receipt expires or is revoked

Example: Agent requesting rates

Agent: request_access({
  identity: { id: "agent_xyz", name: "TalentBot", type: "agent" },
  purpose: "Evaluate fit for a UX audit project for ClientCo",
  scopes: ["read:rates", "read:availability_detail"]
})
→ Pending approval from harness owner
Owner: Grants access (timeboxed, 24 hours)
→ Consent receipt issued: cr_abc123
Agent: read("harness://offer/rates", { receipt: "cr_abc123" })
→ { currency: "AUD", units: [...], rate_rules: [...] }

Open questions

How does the agent authenticate the consent receipt?
Should there be a capability URL pattern instead of explicit receipts?
How do we handle agent-to-agent delegation?