The world's first AI-first company. You will work side by side with superintelligence — the direction the United States has declared for the decade. People and AI agents are co-executors: agents plan and do the work on our own platform, people sign what matters. We hire in Ukraine, the EU, the United States and Japan.
About the role
You work on the agent runtime itself: how a Planning agent decomposes a mission, how Builder and QA agents hand work to each other, how a Knowledge agent retrieves and cites, how guardrails and human signatures gate every action that leaves the sandbox. Stack: Deno/TypeScript on the platform, Python where models need it, MCP for tools, A2A between agents, SQLite/KV persistence, local and hosted LLMs with model fallback.
Working with AI — your co-executors
You build the co-executors everyone else works with — and you work with them yourself. Snaga Code pairs with you in your repo: it drafts, you review; a QA agent runs the evaluation suite before anything reaches a tenant. You decide what an agent may do without a human signature, and you measure whether the agents got better this week.
You are not using AI tools; you are working in a team where AI agents are co-executors. Every role at Snaga has its own pod of agents on the Snaga AI platform: a Planning agent that turns a brief into a plan, Builder agents that do the first pass of the work, a QA agent that checks it, and a Knowledge agent that answers with citations from our documents. Agents plan and execute; the person sets the goal, reviews the result, signs every action that leaves the sandbox, and teaches the agents where they were wrong. Every step is in an immutable audit log, so you always see what the agents did and why.
Our tools, exclusively
All work and all training happen on Snaga's own instruments: the Quark browser, the Snaga AI orchestration platform and the Svitlo agent. Quark delegates work by conversation — to the Snaga Code agent for development and to Svitlo for information research — keeps every account in its own isolated profile, and asks for your explicit approval before anything touches passwords, e-mail, payments or deletions. On the platform your work runs as agent teams (Planning, Builder, QA, Knowledge) in a Plan→Build→QA→Deliver→Learn loop; every action that leaves the sandbox waits for a human signature, every step is in an immutable audit log, and knowledge agents answer with citations. You will be trained on these tools in your first weeks and you will help make them better.
How we work
Remote-first, async by default, with a weekly review. Tooling is ours; equipment is on us.
What you will do
- Design and ship agent behaviours: planning, tool calling, memory, multi-agent handoff
- Build retrieval that answers with citations and measures its own precision
- Write evaluations that turn red before a regression reaches a tenant
- Integrate models and providers behind one adapter with fallback
- Review the Snaga Code agent's output and improve its prompts and tools
- Work with AI agents every day as co-executors: brief them, review their output, sign what matters, and teach them where they were wrong
What we expect
- 4+ years of software engineering, 1+ shipping LLM-based systems in production
- TypeScript (Deno or Node) and Python
- Hands-on with tool-calling agents, RAG, embeddings and evaluation harnesses
- Understanding of MCP or similar tool protocols
- Readiness to work and train exclusively on Snaga's tools: Quark, the Snaga AI platform and the Svitlo agent
- Comfort working alongside AI agents as co-executors: delegating, reviewing their output critically, and taking responsibility for the signed result
- English B2 or higher
Nice to have
- Experience with local inference (Ollama, vLLM) and quantised models
- Contributions to open-source agent tooling
What you get
- Remote, flexible hours
- Equipment and tooling budget
- Paid training and certifications
Skills
- TypeScript
- Deno
- Python
- LLM agents
- RAG
- MCP
- Evaluation
- Snaga AI