Agentic AI Engineer

About the role:

We build enterprise AI agentic solutions that integrate with the business systems teams already use. An agent holds a real conversation with a business user, gathers live context from those systems, reasons over it, recommends a course of action, and — only on explicit human approval — writes changes back to the system of record. For example, an agent can take a user from a plain-language request to a validated, policy-checked action — a draft transaction or record ready for a person to approve — without leaving the flow of a chat.

We build these on a shared platform: a code-first agent runtime, a reusable intelligence layer, and a governed integration surface into the connected systems. Every solution is model-driven, grounded in real system-of-record data, and gated by a human on anything that changes those systems. We are opinionated about three things: the human stays in control, every recommendation is explainable, and the system of record remains the single source of truth.

You’ll be a core builder on the agent-and-platform side — the heart of what we ship. You’ll implement agent logic, the decision loop that drives it, and the governed tool layer that lets an agent read from and (carefully) write to connected systems. This is a backend/agent-heavy role: Python, LLM orchestration, tool/function calling, and enterprise integration — not a frontend role, though you’ll touch the chat conversation surface where the agent lives.

You’ll work within patterns we’ve already established — a mid-level hire builds confidently inside a proven architecture and helps harden it — with a senior architect and PM setting direction. You’ll own features end-to-end: from a technical design handed to you, through implementation, tests and evals, to a reviewed, demoable agent.

Key Responsibilities:

• Build agent behavior on a code-first agent runtime (Python) — implement the multi-stage decision loop (observe → detect → explain → predict → recommend → approve → execute → measure → learn) that turns a user’s ask into a grounded recommendation.
• Design and write prompts that force grounding — the agent cites evidence from live data or declines; no fabricated numbers, entities, or policy.
• Implement the human-approval gate on every write path, so nothing changes a connected system without an explicit, auditable confirmation from the user.
• Build and extend the tool layer (built on the Model Context Protocol) that connects the agent to business systems and data sources — scoped read tools and gated write actions, all under delegated (on-behalf-of) identity, never a broad service principal, never direct database access.
• Map real integration failure modes (throttling, locked or closed records, stalled workflows, auth expiry, validation rejects) to concrete retry/backoff and graceful-degrade behavior — not hand-waved error handling.
• Write the test and eval suite: unit tests tied to acceptance criteria, plus agent evals for task success, grounding/faithfulness (hallucination), approval integrity, and cost/latency against defined SLOs.
• Keep an eye on cost-per-conversation and latency — model-driven orchestration is powerful but not free; you’ll help tune tool consolidation, prompt caching, and model choice to stay inside budget without losing quality.
• Contribute to the shared platform packages (intelligence layer, tool/integration client, observability) so new solutions reuse them instead of reinventing.
• Participate in design and code review, and help keep the codebase clean as the portfolio grows from one solution to many.

What you will work with:

• Python · a code-first agent SDK · Model Context Protocol (MCP) · LLM tool/function calling and prompt engineering · a hosted model platform (BYOK) · enterprise system integration over APIs (REST / OData) · identity and delegated auth (OAuth / OIDC — e.g. Microsoft Entra / MSAL) · a chat surface such as Microsoft Teams or Slack · pytest and agent-eval harnesses · OpenTelemetry-based observability · Git/GitHub, CI.

Must have:

4 to 7 years building production software in Python, with a track record of shipping and maintaining real systems — not just prototypes.

Hands-on experience building LLM-powered features or agents — you’ve worked with an LLM API, done tool/function calling or an agent loop, and written prompts that had to behave reliably in production. You understand why an agent halluci…