Principal AI Engineer

To get the best candidate experience, please consider applying for a maximum of 3 roles within 12 months to ensure you are not duplicating efforts.

Job Category

Software Engineering

Job Details

About Salesforce

Salesforce is the #1 AI CRM, where humans with agents drive customer success together. Here, ambition meets action. Tech meets trust. And innovation isn't a buzzword - it's a way of life. The world of work as we know it is changing and we're looking for Trailblazers who are passionate about bettering business and the world through AI, driving innovation, and keeping Salesforce's core values at the heart of it all.

Ready to level-up your career at the company leading workforce transformation in the agentic era? You're in the right place! Agentforce is the future of AI, and you are the future of Salesforce.

We are seeking a highly skilled AI Platform Engineer to play a pivotal role in building the next generation of our ML/AI platform that doesn't just support ML models, but powers autonomous AI agents at enterprise scale. This role sits at the intersection of platform infrastructure and agent systems engineering. You'll build and maintain the core infrastructure, CI/CD pipelines, and platform services that underpin our machine learning initiatives and go further in designing the harnesses, sandboxes, and evaluation frameworks that let AI agents be developed, tested, and trusted in production.

You'll work on systems that directly impact marketing, sales, service, and product growth verticals across the organization.

This isn't a traditional infrastructure role. You should be comfortable wearing multiple hats of software engineering, agent systems design, and evaluation tooling. We're looking for engineers who think in flywheels: build →evaluate → improve → ship → repeat.

What You'll Do

Agent Harness & Flywheel Engineering

* Design and build agent harness infrastructure: the scaffolding that wraps LLM calls, manages tool use, handles retries, enforces policy, and feeds results back into iterative improvement loops.

* Implement agentic loop patterns with multi-turn reasoning, tool orchestration, memory management, and structured output handling as reusable platform primitives

* Build the agent flywheel: automated pipelines that collect agent traces, surface regressions, route failures to evaluation, and close the loop from production signal back to prompt/model improvement

* Own the end-to-end lifecycle from agent experiment to production deployment, including versioning, rollout controls, and rollback mechanisms

Sandboxing & Safe Execution

* Build sandboxed execution environments for agent tools with isolating code execution, API calls, and file system access so agents can act without unconstrained blast radius

* Design tiered autonomy models: define which actions agents can take automatically, which require human approval, and which are off-limits and enforced at the infrastructure layer

* Implement replay and dry-run capabilities so new agent versions can be tested against real traces before going live

Agent Evaluation, Observability & Optimization

* Implement evaluation frameworks for agent behavior using a combination of vendor , open source or in house built tools - covering task success, tool selection accuracy, trajectory evaluation, hallucination rates, latency, and cost

* Build and maintain eval datasets, golden trace libraries, and regression test suites that run automatically on every agent code change

* Instrument agent traces end-to-end: LLM calls, tool invocations, intermediate reasoning, final outputs - surfaced in Grafana or equivalent observability tooling

* Define and track agent quality metrics over time; own the signal that tells the team whether agents are getting better or worse

* Drive continuous quality, latency, and cost improvements across deployed agents by closing the loop between production traces, evaluations, and agent design. Optimization may be done through a variety of techniques e.g. prompt tuning, tool calling optimizations, context engineering, right-sizing model selection per task and explore distillation or fine-tuning (SFT, DPO, RLHF) on curated trace data to name a few

*…