← All roles

Senior Harness Engineer

Pydantic· Engineering
Apply on company site
Remotefull-timePosted 3mo ago

About the role

Own and develop Pydantic AI Harness, an open-source Python library for building production-grade AI agents. Build reusable agent capabilities, evaluate and improve production agent performance, and contribute to Pydantic AI’s extension APIs.

Responsibilities

  • Own the technical direction and day-to-day development of Pydantic AI Harness, working primarily in Python.
  • Design, build, and maintain reusable agent capabilities such as context management and compaction, memory, planning, skills, sub-agents, steering, sandboxed tool execution, guardrails, monitoring, and recovery from failed or stuck runs.
  • Build with the model as a user of the system: make state, provenance, errors, limits, and changes to its environment legible so it can act on an accurate view of the world.
  • Dogfood the harness in real agents and agentic development workflows, then turn what breaks into better APIs, capabilities, documentation, and defaults.
  • Instrument agents with Pydantic Logfire, analyze production traces, create evaluations and benchmarks, and verify that changes improve task success, reliability, latency, and cost rather than merely looking plausible.
  • Establish reproducible benchmarks against leading agent harnesses and keep pace with a fast-moving field through experiments, competitor research, and close attention to new models and techniques.
  • Share what you learn in public: write up experiments, benchmarks, and design decisions in blog posts and talks that help define what good harness engineering looks like.
  • Work across Pydantic AI and the harness: improve the capabilities and hooks extension APIs in core when a clean harness implementation needs framework support.
  • Maintain a high-quality open source project the way it works in the agentic era: set and hold the quality bar for a codebase that is largely agent-written, talk with production users on Slack, calls, and GitHub to learn what they actually need, iterate with them on the design of proposed capabilities — then drive the implementation with your own agents — and communicate API changes clearly.
  • Ship quickly while preserving trust through strong tests, useful deprecation paths, clear documentation, and evidence from real agent runs.

Qualifications

  • Significant experience building, operating, and improving LLM agents or agent harnesses in production, particularly agents that work autonomously over long-running, complex tasks. You know the difference between an impressive prototype and a system that remains useful under real workloads.
  • Deep understanding of the agentic loop, from Pydantic AI or any comparable agent framework — agent loops, tools, context management, structured outputs, model-provider differences, and the abstractions frameworks use to bring them together. The concepts transfer; framework-specific knowledge can be learned.
  • Solid Python engineering judgment. Much of our code is written by coding agents under direction, so what matters most is knowing what good looks like — API design, type safety, tests, backward compatibility — and holding that bar while directing and reviewing agent-written code.
  • Experience instrumenting production agents with an LLM observability platform, diagnosing their behavior from traces, and using that evidence to improve them. You can turn a vague instruction like “make the agent better” into hypotheses, experiments, and measurable results.
  • Hands-on experience with the systems around a production agent: sandboxed code or tool execution, permissions and safety boundaries, long-running state, and multi-agent orchestration.
  • Experience connecting agents to external tools and services through MCP or comparable tool and context protocols.
  • End-to-end ownership. You can identify the important problem, research the landscape, design the right abstraction, implement it, document it, and validate it with users without waiting for a detailed specification.
  • Good judgment about the interaction between model behavior and software design. You understand that prompts, tool schemas, errors, context management, and execution semantics all shape what an agent can reliably do.
  • The instincts it takes to own an open source repo: an appreciation for the weight of public API decisions, backward compatibility and useful deprecation paths, good documentation, a taste for reusable abstractions that let people do more rather than locking them into one way of doing things, and working directly with a technical community in public. Prior maintainership of a popular project is a plus, not a requirement.
  • Fluency driving coding agents as a primary way of producing software — knowing what to ask for, how to specify a design, and how to review the result — while retaining responsibility for architecture, correctness, and quality.
  • At least 5 years of software engineering experience.
  • Experience designing agent evaluation suites or working with benchmarks such as Terminal-Bench, SWE-bench, or similar task-based evaluations.
  • Contributions to open source agent tooling, Python developer tooling, SDKs, or frameworks.
  • Experience maintaining an open source project or helping its contributor community grow.
  • Familiarity with OpenTelemetry and Pydantic Logfire specifically.
  • Experience with ACP or other protocols for connecting agents to user interfaces.
  • Live and work in a timezone between PT (UTC-8) and CET (UTC+1)
  • Able to travel to the EU, UK, and US up to 4 times a year to join our off-sites

Skills

Required:Production LLM agent developmentAgent frameworks and agentic loopsPythonPython engineering judgmentLLM observabilityProduction agent systemsMCP and tool protocolsEnd-to-end ownershipModel behavior and software designOpen source project ownershipCoding agentsSoftware engineering
Nice to have:Agent evaluation and benchmarksOpen source contributionsOpen source maintenanceOpenTelemetry and Pydantic LogfireACP

Benefits

  • Competitive salary and stock options
  • Work from anywhere within our timezone range - no office requirements
  • Join a multi-cultural team of 8+ nationalities
  • Direct influence on tools used by millions of developers worldwide
  • Regular opportunities for learning and professional development
  • Connect with the team at our regular international off-sites
  • Comprehensive health coverage for you and your dependents
  • Work when you're most productive
  • Budget for your home office setup
  • flexible working hours and 33 days PTO no matter where you live (including public holidays, which you can choose to take or not)
Last updated Nov 14, 2026. Interested? Apply on the company site.

About Pydantic

End-to-end AI engineering stack

Pydantic is the AI engineering stack for teams building with AI. Observe with Logfire, evaluate with Evals, build type-safe agents with Pydantic AI, and track costs with AI Gateway.

OSSAgent FrameworkGraphsMCPData ValidationType HintsEvalsOpenTelemetry

More open roles at Pydantic