Architecture & developer documentation

Decoupled agent orchestration

(not)Human Resources sits as a secure, zero-trust infrastructure proxy between your application logic, your internal data stores, and upstream frontier models. Instead of embedding brittle interceptors, rate-limiters, and logging wrappers inside your code, you route agent intents and tool executions through the harness — and every agent runs on a verifiable NEUS identity with an enforceable delegation.

Application logic · LangChain / AutoGen
↓ signed payload
(!) The Harness
MCP Router
Schema enforcer
FinOps Guardrails
Cache · budget circuit
Immutable Ledger
Crypto-signed traces
NEUS Trust
Identity · delegation
Internal DBs · via MCP
Upstream LLM · proxy route

Native Model Context Protocol integration

The harness builds natively on MCP. By formalizing the connection between applications and data sources, it decouples data access from model intelligence.

Harness as an MCP gateway

Your internal tools, databases, and microservices expose standard MCP server endpoints. The harness is the secure client gateway — managing authentication, connection pooling, and payload scrubbing. Identity and delegation ride the NEUS MCP server (mcp.neus.network/mcp).

Context isolation

Upstream models never ingest naked, unvetted environment context. The harness forces all tool discovery and resource sampling to match explicit JSON-schema invariants before they reach the model.

The trust layer

Every agent runs on NEUS

Trust isn't a feature you configure after the fact — it's the substrate. When you hire an agent, the harness registers it as a real NEUS Trusted Agent: an agent-identity and agent-delegation proof pair written to your NEUS profile. The delegation is what the harness enforces before any tool executes.

Verifiable identity

Every hire calls neus_agent_create — an agent-identity proof written to your NEUS profile, on the agent's own dedicated wallet. Not a label; a credential that verifies and revokes independently of this platform.

Scoped delegation

The agent-delegation proof carries the exact allow/deny action lists, a spend cap, and an expiry. The harness checks it before every tool call — denied actions never execute; they wait for one-click human approval at the gate.

Tamper-evident receipts

Every decision — allowed and denied — is hash-chained and signed. The trail verifies offline, so your audit doesn't depend on our uptime or a single vendor's logs.

Real revocation

Offboarding owner-revokes the identity and delegation proofs on NEUS. The agent's authority dies at the source, not just at our gate.

Because identity, permissions, and audit live on NEUS — not baked into your app against one vendor — you swap frontier models with a config line while your security posture and human-in-the-loop policies stay intact.

Core infrastructure

Circuit breakers & budget enforcement

Every session declares a metadata policy with a maximum compute spend. The state engine evaluates cumulative token consumption and latency per turn; a recursive loop or unhandled exception that breaks your boundaries drops the connection at the socket level.

Semantic caching engine

Before forwarding context to upstream providers, the harness runs an async semantic lookup against a low-latency vector cache. Exact and high-similarity hits resolve locally in sub-10ms, slashing redundant token cost by 40–80%.

API design & integration

Session initialization

Send an explicit session policy to spin up a containerized execution harness.

POST /v1/sessions/initialize

{
  "session_id": "agent_wf_99281a",
  "application_id": "crm_outbound_auto",
  "policy": {
    "max_budget_usd": 2.50,
    "circuit_breaker": {
      "max_recursive_depth": 5,
      "action": "terminate_and_alert"
    },
    "routing_strategy": "dynamic_cost_optimized",
    "allowed_mcp_servers": [
      "salesforce-reader-production",
      "slack-notifications-internal"
    ]
  },
  "initial_context": { "user_id": "usr_4021", "scope": "read_only" }
}
200 OK

{
  "status": "initialized",
  "harness_token": "nhr_sec_tok_8f3d192c73be",
  "endpoints": {
    "mcp_gateway_url": "https://gateway.nothumanresources.xyz/v1/mcp/agent_wf_99281a",
    "llm_proxy_url":   "https://gateway.nothumanresources.xyz/v1/chat/completions"
  }
}

Executing a monitored tool call

When the model issues a tool call, the application routes the intent through the proxy to guarantee schema compliance and check the NEUS delegation.

POST /v1/mcp/execute

{
  "harness_token": "nhr_sec_tok_8f3d192c73be",
  "mcp_server": "salesforce-reader-production",
  "tool_intent": {
    "name": "update_lead_status",
    "arguments": {
      "lead_id": "00Q8W00000TvK1u",
      "status": "Qualified",
      "override_protection": false
    }
  }
}
⚠ Zero-trust enforcement: if the tool intent deviates from your JSON schema, targets an unauthorized collection, or falls outside the agent's NEUS delegation, the harness blocks the write and logs a telemetry alert with an immutable cryptographic signature.

Deploy an agent on a verifiable harness.

Deploy the harness