Member of Technical Staff: Agent Runtime

Member of Technical Staff: Agent Runtime (Ego) Location: San Francisco, CA (hybrid) · Reports to: Founding team

About the role

Ego AI is a YC-backed applied AI research lab building the behavioral infrastructure for AI companions and agents. We work at the intersection of real-time conversational instincts, memory, and persistent identity; the layer that makes AI feel genuinely alive. We're a small, fast-moving team and we're defining a new category of human-AI interaction.

What you'll do

Own the agent harness. Design and maintain our core agentic loop (plan → tool call → observe → repeat → finalize) as a small, legible state machine. Errors are first-class: malformed tool calls, hallucinated tool names, and throwing tools get fed back to the model. They never crash the loop.

Make execution durable. Build checkpointed, resumable sessions backed by a database: step logs with pending/committed status, idempotency keys on side-effecting tools, and a clear account of the at-least-once vs exactly-once boundary. A crash between the LLM call and the tool call, or mid-compaction, must never corrupt a session or double-fire a POST.

Solve long-horizon context. Own our compaction strategy: pinned goals, running summaries, sliding windows of recent turns, and retrieval over older state. A 30+-step task should finish aimed at the original goal, within budget.

Design for extensibility. Ship a tool registry (name + JSON schema + handler) that teammates and partners can extend without touching the loop. Validate schemas before handlers run. Keep model adapters swappable in one place.

Ship deployable systems. Deliver one-command local bring-up, persistence behind an interface (Postgres ↔ SQLite ↔ Cloudflare storage), clean HTTP APIs, and secrets hygiene. Build per-user durable agent instances that wake on triggers (cron, webhooks, inbound events). Cloudflare Durable Objects experience is a real plus.

Support the team. Unblock product engineers building on the harness, pair on extensions, and review agent-adjacent designs.

Projects this hire will own

Productionize the agent runtime ("Ronin core"). Take our harness from working prototype to the shared runtime every Ego product sits on: durable step log, budget enforcement, compaction, tool registry, and observability/tracing.

Per-user durable personal agents. Each user gets a long-lived agent instance with its own memory, triggers, and budget. It wakes on webhooks or cron, survives restarts, and stays isolated per tenant. Likely on Cloudflare Durable Objects or an equivalent single-writer model.

Evaluation and reliability harness. Build repeatable crash tests (kill -9 mid-task and mid-compaction), race tests on concurrent session writes, cheap long-horizon stubs that exercise compaction without burning tokens, and per-step tracing.

What we're looking for

Must have

• 3+ years building backend or infrastructure systems, with real production ownership of stateful services or relevant projects that they have worked on.
• Hands-on experience building LLM agent systems: tool-calling loops, defensive handling of model output, and step/token/cost budgets.
• Deep grasp of durable execution: checkpointing, idempotency, outbox/step-log patterns, and the ability to reason precisely about failure windows (what if it dies after the side effect but before it is recorded?).
• A deliberate context-management strategy for long-horizon tasks, and the ability to defend its trade-offs rather than only describe it.
• Strong API and data-model design; persistence behind clean seams; comfort being judged on docker compose up working cold from a README.
• Clear written communication. Architecture docs a reviewer understands before reading the code. Honest scoping (what you cut and why).

Nice to have

• Cloudflare Workers / Durable Objects / D1 in production.
• Multi-model experience (DeepSeek, Claude, open-weight models) and eval tooling.
• Long-term memory systems for agents (personalization layers, retrieval over older state, hierarchical summaries).
• Concurrency safety on shared session state; streaming output; structured per-step observability.

Use AI tools freely. We do. You will walk us through the design and defend every trade-off, so own each decision in the repo.

Interview Process

How we hire

1. Intro chat (15-30 min).
2. Take-home: build a durable tool-using agent.
3. Live pairing (45 min): extend your own submission. Add a tool, make a race safe, and walk through every point a kill -9 could lose or duplicate work.
4. Behavioral interview
5. Founder conversation + references.