Founding Engineer, Agentic Systems

Champ AI•San Francisco, CA

2d•Onsite

About The Position

Champ AI is building a multimodal work-agent orchestration platform that helps ops/support/compliance teams automate end-to-end workflows. We're building agentic systems that can reliably take actions across tools, handle real-world edge cases, and continuously improve with evaluations and feedback loops. The Role We're looking for a Founding Engineer (Agentic Systems) to own core pieces of our agent runtime and developer/product surface area. You'll build the systems that let agents operate safely, deterministically, and measurably in production: memory + context management, tool integration, sandbox execution, data syncing, evals, and AI-native UX. This is a hands-on role where you'll ship to production quickly, work directly with design partners, and help define what "good" looks like for enterprise-grade agents.

Requirements

Shipped AI agents into production (internal or external), or Built meaningful open-source contributions in agent frameworks, eval tooling, RAG/memory tooling, browser automation, or similar.
Experience with LLM tool-use, structured outputs, function calling, and multi-step workflows.
Experience with context engineering: retrieval strategies, chunking, reranking, summarization, memory write/read policies.
Experience with systems thinking: state machines, retries, idempotency, failure modes, and "what happens at 3am."
Experience with integrations: OAuth, scopes, token refresh, pagination, incremental sync, webhooks, rate limiting.
Experience with sandboxed execution or secure-by-default infra patterns (containers, ephemeral environments, secrets).
Experience with observability: traces, metrics, logs; building "explainable runs" for humans.
Experience with evaluation approaches for non-deterministic systems; confidence scoring; regression testing.

Nice To Haves

Built AI-native UI surfaces (not just APIs): agent run views, trace explorers, approval UIs, etc.
Worked with enterprise requirements: SOC2 posture, auditability, access controls, tenant isolation.
Can move between research-y prototyping and production-grade engineering without getting stuck in either.

Responsibilities

Agent runtime + orchestration
Agent loop design (planning → tool-use → verification → recovery) with strong guardrails.
Context assembly pipelines: retrieval + compression + summarization + "state" that survives long workflows.
Memory management: short-term working memory, long-term memory, user/org/project memory, and safe write policies.
Multi-agent patterns: delegation, handoffs, coordinator/worker setups, and concurrency.
Tool definition frameworks: typed schemas, validation, retries, idempotency, rate limits, and observability.
Connectors + data syncing: SaaS APIs, webhooks, polling strategies, incremental sync, conflict resolution.
Browser automation / computer-use flows (auth, session handling, DOM variability, screenshots, network traces).
Sandbox + execution Secure execution environments for "agent writes code / runs scripts / transforms data."
Permissions, isolation, secret management, and audit trails.
Deterministic replays where possible; safe "dry run" modes; blast-radius controls.
Evaluation harnesses for tool-use correctness, workflow completion, policy compliance, and regression detection.
Golden tasks + synthetic tasks + real production traces; offline + online metrics.
Experimentation frameworks (prompt/model/tool changes), versioning, and rollbacks.
Human-in-the-loop review flows: sampling, labeling, adjudication, continuous improvement loops.
Interfaces that make agents understandable and controllable: traces, state, "why it did that," and editable plans.
UX patterns for approvals, step-through execution, partial automation, and exception handling.
Customer-facing debugability: audit logs, run history, data provenance.