Responsibilities :

- Build and operate the agentic loop : trigger - orchestration - agent execution - output to JIRA - human accept/reject - next agent, across design, coding, review, and testing agents.

- Implement model routing and retry logic across a provider-agnostic model layer (e.g., Claude via AWS Bedrock, self-hosted or alternative models as cost/sovereignty hedges), including business-continuity fallback if a given provider becomes unavailable.

- Own token cost control and context window management - per-agent and per-run budgets, circuit breakers that halt runaway execution, and cost observability tied back to JIRA.

- Stand up and maintain observability, alerting, and monitoring across the agent fleet (e.g., Langfuse or equivalent), so agent health, cost, and quality are visible in real time.

- Implement agent governance and safety guardrails : deterministic pre/post hooks gating every LLM call, kill switches, prompt injection prevention and mitigation, and audit logging.

- Integrate the harness with JIRA as the system of record and other business systems as needed, ensuring every agent action, decision, and human override is tracked with no side channels.

- Pair directly with client engineers throughout - this is capability transfer, not black-box delivery. You'll document, demo, and hand over as you build.

- Work in outcome-based delivery stages (spike - architecture sign-off - build - pilot) with gated milestones tied to working software demos, not fixed artifact checklists.

- Participate actively in team discussion and design decisions - this team expects engineers to challenge ideas constructively and speak up, not defer silently.

Must-Have Experience :

- Hands-on production experience building agentic systems (not tutorial-level or personal-project experience). Candidates should be able to speak concretely about systems they've shipped.

- Practical experience with agentic frameworks such as LangChain, LangGraph, or equivalent orchestration frameworks.

- Experience with LLM orchestration and model routing across multiple providers/models, including fallback and retry design.

- Working knowledge of agent governance : guardrails, human-in-the-loop approval flows, kill switches, and audit trails.

- Practical understanding of prompt injection risks and mitigation techniques.

- Experience with token cost management and context window/memory handling at production scale - this is a named governance requirement for the engagement, not a nice-to-have.

- Strong Python (or equivalent) engineering background, comfortable working in AWS environments (Bedrock/AgentCore exposure a strong plus).

- Experience with observability/monitoring tooling for distributed or agentic systems (e.g., Langfuse, Datadog, or equivalent).

- Comfortable working with JIRA/Atlassian APIs or similar ticketing-system-of-record integrations.

Nice to Have :

- Direct experience with AWS Bedrock AgentCore, Temporal (or similar workflow orchestration), or LiteLLM-style model gateways.

- Exposure to Cursor or other AI-native IDEs in a production engineering context.

- Experience with self-hosted open-weight models (e.g., DeepSeek, GLM) as cost or sovereignty hedges alongside commercial APIs.

- Financial services or other regulated-industry background.

- Familiarity with Claude Code, Claude Cowork, or Claude Skills.