OpsPilot - Spec-Driven AI Operations Agent

An operations copilot that turns plain-language tasks into planned, verified, multi-step work - with human approval for risky actions, an audit ledger and built-in evaluations.

Node.jsMCP (JSON-RPC)HTTP APIEvaluation framework354 automated tests

5/5

Finance workflow scenarios passed

3/3

Policy evaluation suite passed

1/1

Runtime evaluation passed

NOMINAL

Assurance risk level

Runtime pipeline

Parse → Plan → Authorize → Execute → Verify → Report

Assurance console (real output)

OPS PILOT / ASSURANCE CONSOLE
════════════════════════════════════════════════════
PASSED  ·  RISK NOMINAL  ·  PROFILE finance-reference
Run: hf-demo-001    Generated: 2026-11-06T20:12:19.380Z
────────────────────────────────────────────────────
EVALUATION         3/3 passed (100.0%)
POLICY GATE        100.0%
EVIDENCE BINDING   100.0%
EVALUATOR          100.0%
────────────────────────────────────────────────────
DECISION DISTRIBUTION
allowed            1
requires_approval  1
blocked            1
────────────────────────────────────────────────────
FAILED GATES
None — all evaluation gates passed

Finance reference workflow - evaluation scenarios

Scenario	Expected	Decision	Status	Policy	Evidence	Audit	No unsafe prep	Result
`workflow-allowed-payment`	allowed	allowed	prepared	PASS	PASS	PASS	PASS	PASS
`workflow-blocked-vendor`	blocked	blocked	blocked	PASS	PASS	PASS	PASS	PASS
`workflow-pending-approval`	requires_approval	requires_approval	pending	PASS	PASS	PASS	PASS	PASS
`workflow-approved-payment`	requires_approval	requires_approval	prepared	PASS	PASS	PASS	PASS	PASS
`workflow-rejected-payment`	requires_approval	requires_approval	rejected	PASS	PASS	PASS	PASS	PASS

What is inside

Spec-first agent

The agent's role, principles and task loop are a versioned contract (agent-instructions.md), so behaviour is reviewable - not hidden in prompts.

Human-in-the-loop

Irreversible or outward-facing steps pause for approval; blocked vendors and over-threshold payments never execute unchecked.

MCP + HTTP boundary

Finance tools exposed through an MCP JSON-RPC adapter and a dependency-free HTTP adapter with timeouts and structured errors.

LLM planner, deterministic fallback

Gemini, OpenAI or Claude propose the plan; it is validated against the plan contract and falls back to the deterministic planner on any error. The model proposes, policy decides.

Prompt-injection defense (OWASP LLM01)

Tasks, tool arguments and HTTP input are screened for instruction overrides, secret exfiltration, tool smuggling and hidden characters; risky input is blocked or sent to a named reviewer.

Observability

Prometheus /metrics (runs, stage durations, approvals, blocks, fallbacks) and OpenTelemetry-style traces with one span per pipeline stage.

Idempotency keys

The same request key never prepares a payment twice; a changed payload with the same key is rejected with a clear conflict error.

Evaluation control plane

Every run carries a schema version, run ID and per-gate results (policy, evidence, approval, audit) - it fails CI if any gate regresses.

The numbers and console above were produced by running the OpsPilot runtime from the public repository (2026-11-06). Simulation-only: no real payments, companies or personal data. Source: github.com/furqunali/ops-pilot-agent · Author: Furqan Ali - Senior AI Engineer · Forward Deployed Engineer