Job Description
The Mission: The enterprise software landscape is fracturing. We are transitioning from "Systems of Record" to "Systems of Action." Zendesk is leading this shift with the Resolution Platform, aiming to be the first CX company to secure $1B in AI-driven revenue. To achieve this, we must solve the "Orchestration Trap." We are not just building features; we are building the high-performance runtime that allows fleets of AI Agents to Perceive, Reason, and Act.
The Opportunity: You will operate at the cutting edge of large language models, planning algorithms, and multi-agent coordination. By designing advanced memory structures and autonomous learning mechanisms, you will bridge the gap between offline model capabilities and dynamic, real-world task execution.
What we have built
We run production AI agents that autonomously resolve customer service tickets across 100,000+ Zendesk accounts. These agents take a customer issue, plan a multi-step resolution, execute real actions (refunds, order modifications, escalations) through live APIs, and close the ticket without a human in the loop.
The agent core uses a proprietary iterative architecture: the agent decomposes goals into plans, pulls reusable skills from a registry, executes, evaluates the outcome, and refines. Each iteration feeds back into the next attempt. We have a working self-learning mechanism where successful resolution patterns are synthesized into new skills and fed back into the registry, so the system improves from its own execution history.
On multi-step tool-use benchmarks (GAIA-class), our agents perform at parity with the best published results. Our internal evaluation suite runs 158+ scenario-based tests from real Zendesk tickets, scored continuously through Braintrust with regression detection on every deploy.
What we need help with
Pushing the architecture further. The iterative planner works, but there are open questions we have not solved yet: how to handle plan decomposition when the goal is ambiguous, how to manage interference between memory tiers under concurrent sessions, how to make skill acquisition more selective (the agent acquires skills too eagerly today), and how to design multi-agent delegation patterns where one agent hands off subtasks to specialized agents via A2A (the Agent-to-Agent protocol).
Domain-specialized agent models. We are building toward training our own models, specialized for customer service resolution via RL on production trajectories. The data pipeline is already being instrumented (resolution outcomes, escalation patterns, user satisfaction signals). The next step is the RL training infrastructure itself: reward curricula, rollout systems, and the feedback loops that turn a capable base model into a specialist that matches or beats frontier models on our task distribution at significantly lower inference cost. This is a 6-12 month build, and we need someone who can own both the science and the systems.
Hardening evaluation. We run 158+ scenario evals continuously with regression detection, but multi-turn evaluation and automated trajectory analysis (pinpointing where reasoning diverged) are still early. We need quality gates that block deploys when agent performance drops, and we need them integrated into CI, not run as an afterthought.
Guardrails at enterprise scale. The threat surface for autonomous agents includes tool misuse, cascading action chains, prompt injection, and hallucination loops that burn tokens before anyone notices. We need multi-layered defenses with supervisor patterns, capabilities-based access control, and output validation that works across thousands of concurrent sessions without adding meaningful latency.