Agent Performance & Internal Reliability

Phosphor builds software with people and AI agents working together. Our Software Factory Manager owns two connected responsibilities: improving the speed and correctness of agent work, and keeping the internal infrastructure behind that work dependable.

This is a hands-on individual contributor role, initially with no direct reports. You’ll investigate agent sessions, develop tools, change code, and operate services.

Improve agent speed and correctness

Understand how individual agents perform in our codebase: where they spend time, what they misunderstand, why they repeat work, and whether their results satisfy the task.

• Develop tools to inspect and compare agent sessions, including tool calls, verification results, failures, and time spent waiting.

• Identify recurring causes of slow or incorrect work, such as missing context, confusing interfaces, poor diagnostics, or ineffective instructions.

• Improve agent tools, feedback, and working environments to address those causes.

• Evaluate changes against representative tasks, checking both time to a correct result and the amount of human intervention required.

• Turn individual session failures into improvements that benefit future sessions.

Keep the factory running

Maintain the internal developer tools, inference providers and connections, and agent machines that the team relies on.

• Monitor service health and detect failures before they interrupt substantial work.

• Diagnose outages, restore service, and prevent recurring incidents.

• Maintain deployments, configuration, access, upgrades, and recovery procedures.

• Manage machine capacity, storage, and resource contention so agents can work reliably.

• Make failures understandable: distinguish provider problems, infrastructure failures, tool defects, and errors in the software being developed.

• Keep operational documentation and ownership clear enough that recovery does not depend on one person.

What success looks like

Agents reach correct, verified results faster, with less repeated work and less human rescue. We can explain why a session performed well or poorly, and demonstrate whether a change improved it.

The services and machines behind those agents remain available and predictable. When something fails, we detect it quickly, recover effectively, and reduce the chance of recurrence.

We’ll establish baselines and evaluate progress through representative task completion, correctness, rework, human intervention, service availability, recovery time, and infrastructure-related delays.

You’ll work closely with engineering to choose meaningful tasks and correctness criteria, prioritize improvements, and turn new tooling into dependable everyday infrastructure.