Production AI Engineer - Build Agent Tools & Workflows

The Consensus

San Francisco, Northern (CA, KY)

Hybrid

USD 179,000 - 240,000

Full time

14 days+

Get more replies from employers

Send a job-specific resume in minutes.

Job summary

The Consensus is building AI agents with security and governance at the core, shipping inside Fortune 100 environments. We’re hiring an applied-AI engineer to design and ship agentic tools and workflows that can act across real systems with proper authorization and traceability.

You’ll own tool harnesses, design multi-tool workflows, and run model-aware experiments across leading AI models. If you love turning bold ideas into production-grade systems and shaping the tooling platform for futures

Qualifications

5+ years software engineering experience, with at least 2 years shipping production ML or applied-AI systems.
Strong Python.
LLM application depth — prompting, retrieval, tool use, agent design.
Experience designing or composing multi-tool / multi-agent workflows that produced real outcomes.
You've built evals at scale — not I ran a benchmark once, but a measurement system real engineering decisions were made against.
Statistics fluency — significance, confidence intervals, A/B test design.
Comfort across multiple frontier models and reasoning about their behavioral differences.
A do-er, not a researcher-in-residence. You'd rather ship a working v0.5 next week than a polished v2.0 next quarter.
Comfort with ambiguity — early team, narrow charter that will expand. You make good decisions with incomplete data.
An insatiable desire to ship.

Responsibilities

Design and ship agentic tools that go beyond deterministic API wrappers — and define the patterns the rest of the Tools team will use to build more.
Build the agent harness that automates tool creation — take a vendor's API, produce a high-quality toolkit end-to-end, keep humans in the loop only where humans add real value.
Design workflows that compose tools into higher-level abstractions customers can actually point at outcomes ("triage this inbox," "close out this account") rather than individual API calls.
Bring applied-ML rigor to tool design — evals, model-aware iteration, retrieval, tool description tuning, response shaping. Make decisions defensible with data.
Run model-aware experiments across Claude, GPT, Gemini — agentic tool behavior diverges across models in ways nobody else is studying, and we should.
Set the technical bar for what "good tool-building" looks like as the team scales — your patterns get inherited by every toolkit author after you.
Contribute back to the MCP and agent ecosystem where the conversation about agentic tools is forming.