ENTITY BIND + OPENAI AGENTS

Resolve a company, read the decision receipt, and continue when a customer
provides missing information. Two entry points share the same identity engine:

1. Bring your own agent: download /api/openai-agents/example.py.
2. Investigate a review case: open /assistant on this deployment.

The integration is implemented. Live availability depends on this deployment's
configuration and your credentials. The managed OpenAI Agents API is a public
beta. Its sessions are stored in the US and do not support Zero Data Retention.

RUN THE EXAMPLE

Requirements: Python 3.12+, httpx, an OpenAI key with Agents/model access, and an
Entity Bind resolve credential. Your plan determines available research depth.
Use the public HTTPS MCP URL supplied by your deployment, not a documentation URL.

Download the Python file above into an isolated virtual environment, then:

  python -m pip install 'httpx>=0.27,<1'

Configure these values securely in the environment. Never put keys in case text:

  OPENAI_API_KEY
  ENTITYBIND_API_KEY
  ENTITYBIND_MCP_URL
  OPENAI_AGENT_MODEL              optional; default gpt-6.1-sol

Save a labelled test company to case.txt. Run:

  python entitybind_agent.py run --message-file case.txt --state case-session.json --wait 60
  python entitybind_agent.py status --state case-session.json
  python entitybind_agent.py resume --message-file clarification.txt --state case-session.json --wait 60
  python entitybind_agent.py cancel --state case-session.json
  python entitybind_agent.py delete --state case-session.json

Keep the session-state file and exported reports outside version control. The
state file holds the session ID and previous-turn cursor, without credentials.
The session retains company information at OpenAI and a scoped MCP credential.

On a timeout, retrieve status before creating another session or sending another
message. A disconnected client does not stop remote work. The CLI never retries
mutations automatically. Read --help for exit states and bounded polling.

READ THE REPORT

session_status and turn_status describe agent execution. They do not establish a
correct identity. Inspect failed_tool_calls even after a completed turn.

resolution_decisions contains actual backend decisions; receipts contains the
evidence retrieved with get_decision. next_step indicates whether to request a
missing identifier, review the decision receipt, or inspect incomplete evidence.
An unresolved company remains unresolved. A newly created local entity is not
independent proof of a real-world company. The agent's wording cannot override
Entity Bind's decision, confidence basis or receipt.

No review-submit, CRM, payment or supplier-creation tool is available to this
reference. Resolution may itself create Entity Bind entities, receipts and review
items and consume credits. Identity does not establish bank-account ownership,
payment authority, sanctions clearance or beneficial ownership.

USE THE INVESTIGATION WORKSPACE

Open /assistant and connect with an issued resolve key. Select an unresolved
record and inspect its source evidence. Starting the investigation requires
explicit acknowledgement of the OpenAI data flow. Add a registration number,
clarify an address or ask a follow-up; reopen saved work later with the same key.
The workspace shows agent status separately from the company decision.

A changed or revoked credential may require deletion and creation of a new
investigation. Stop current work to cancel an active turn. Delete investigation
requests remote session deletion; original Entity Bind records and receipts keep
their existing retention policy. OpenAI physical cleanup can finish asynchronously.

The assistant prepares a case. Human verdicts continue through the separate
review credential and existing review API. The public shell contains no tenant
data, has no analytics, and saves no key in browser storage.

AVAILABILITY AND OPERATIONS

GET /v1/investigations/capabilities (resolve credential required) reports whether
this deployment has enabled the assistant. It is disabled by default. An operator
must configure a permitted OpenAI project, public authenticated MCP URL, explicit
data-transfer setting, durable database and shared rate limiting before launch.

OpenAI inference is additional to Entity Bind credits. Session/follow-up/rate
limits are not a dollar cap. Use a dedicated OpenAI project with billing controls
and monitor total cost before enabling customer access. Never promise that a
model instruction limits tool calls, or that new orchestration improves accuracy.

OFFICIAL OPENAI DOCUMENTATION

https://developers.openai.com/api/docs/guides/agents-api/overview
https://developers.openai.com/api/docs/guides/agents-api/tools/mcp
https://developers.openai.com/api/docs/guides/agents-api/sessions

ENTITY BIND REFERENCE

/api/docs                 API documentation
/api/openapi.json         Current deployed REST schemas
/api/mcp/tools.json       Current MCP tools and input shapes
/api/operations           Deployment operations
/assistant                Customer investigation workspace