Senior Product Engineer, Agent Systems

Agents that generate plausible code from priors are easy. Agents that modernize a regulated enterprise's estate and can prove the new behavior matches the old, are the hard problem. That's the problem this team owns.

As Senior Product Engineer on Aedeon's Agent Systems team, you'll build the agents that do the actual modernization work. Aedeon's product surface is the agent fleet itself: parsers that read legacy estates, extractors that lift business rules out of code, synthesizers that stand up agentic workflows, verifiers that produce behavior-equivalence proof. You'll design and ship those agents. This is product engineering, not services delivery. The agents you build are the product. They run against the Code Intelligence Graph, a persistent semantic representation of the customer's estate and operate under governed autonomy: shadow, supervised, and autonomous stages. You'll write Python every day, orchestrate frontier foundation models through the Aedeon Decision Model, and ship into the same release discipline the rest of the product team holds.

You don't need prior Bedrock or LangChain experience. You need to be a strong Python engineer who treats agent frameworks as something to learn deeply, not a buzzword. The customers are real enterprises in regulated industries, with audit chains that must stay intact. We're looking for someone who writes Python they'd be proud to read in a year, treats test cases as a first-class deliverable, and uses AI tools to do the job better and faster.

What will you do?

Agent systems development

• Build and ship specialized agents in the Aedeon fleet: parsers, business-rule extractors, dependency mappers, test synthesizers, behavior replayers, and the orchestration that wires them together.
• Design agents that operate against the Code Intelligence Graph rather than generating from priors, so every agent action traces to a line of code.
• Implement governed autonomy shadow, supervised, and autonomous stages with human-in-the-loop controls at every stage.
• Orchestrate frontier foundation models (Anthropic, OpenAI, Google) through the Aedeon Decision Model: the right model for each task, against the right slice of the graph, under explicit governance constraints.

Release ownership

• Own the delivery of your assigned agents from prototype through deployment and post-release validation.
• Commit to sprint and release deadlines. Raise blockers and risks with enough lead time to mitigate them.
• Coordinate with platform and DevOps engineers to keep deployment pipelines clean and repeatable.

Verification and quality

• Practice test-driven development. Write the tests for the agent's contract, governance constraints, and equivalence checks before the agent code that satisfies them.
• Build behavior-equivalence verification into every agent: dual-run tests, output diffing, equivalence certificates against production traffic.
• Write detailed test cases before deployment functional flows, edge cases, regression scenarios, integration touchpoints.
• Use AI tools to generate qualitative test coverage at scale, and validate it for completeness.
• Maintain automated test suites (unit, integration, E2E) and integrate them into the CI/CD pipeline.

Collaboration and documentation

• Write clear, maintainable Python with adequate documentation. Review pull requests thoroughly and give constructive feedback.
• Document agent contracts, prompt structures, decision logic, and verification approaches in Confluence or equivalent.
• Share patterns and testing practices across the team. Hold the bar on release discipline and quality.

What are we looking for?

Core Python and systems

• Strong Python (4+ years production): async (asyncio), performance optimization, idiomatic code.
• A solid grasp of distributed systems concepts: state machines, retries, idempotency, eventual consistency.
• Experience integrating with LLM APIs (Anthropic, OpenAI, or similar) from production code: streaming, function calling, structured output, retries, prompt management.
• Comfort with FastAPI or equivalent async web frameworks.

Cloud and infrastructure

• Working knowledge of AWS services: EKS, ECS Fargate, S3, DynamoDB, Lambda, Secrets Manager, CloudWatch.
• Experience with Docker and Kubernetes: writing Dockerfiles, Helm charts, and Kubernetes manifests.
• Comfort with CI/CD pipelines, GitHub Actions preferred.
• Comfort navigating multi-account AWS environments (dev, uat, prod).

Testing and automation

• Test-driven development as a discipline. Tests written before the code, not after.
• Hands-on experience with pytest, integration testing, and E2E testing.
• The ability to design behavior-verification harnesses: dual-run, output comparison, equivalence proof.
• Experience using AI tools (Claude, Copilot, LLM-based test generators) to accelerate and improve test case quality.
• Experience integrating automated tests into CI/CD pipelines.

Engineering mindset

• Test-first, release-disciplined, ownership-driven.
• Strong written and spoken English. You'll be in product reviews and customer-impacting design discussions.
• Available to work with US business-hour overlap from India.

You'll be preferred if

• Agentic AI frameworks: AWS Bedrock AgentCore, Strands SDK, LangChain, or similar.
• Temporal.io or other workflow orchestration engines.
• Graph databases (Neo4j) and the Cypher query language.
• Amazon OpenSearch Service or Elasticsearch.
• Java exposure for working with the Java Analyzer component.
• Prior B2B SaaS product work with a structured release process.
• Domain exposure to enterprise modernization: mainframe, SAP, Oracle, or large Java / .NET estates.

How we work?

Aedeon is product-led. You ship agents into a release cadence, not into billable hours and what you ship gets exercised hard, because Mactores deploys Aedeon across its engagements as the platform's flagship deployment partner. That loop is what makes the job compound: field delivery surfaces the edge cases, the platform roadmap absorbs them, and your agents get better against real enterprise estates, not benchmarks. Every claim an agent makes has to trace to a line of code, and every behavior change has to carry proof. If that standard sounds like the interesting part rather than the annoying part, you'll fit here.