Important: This role is open to candidates based in LATAM, Africa, and Eastern Europe. Please note that as this role supports U.S.-based clients, candidates must be available to work during U.S. business hours aligned with the client’s time zone.
You build production agent systems that run themselves. Not prompts, loops. Not a single assistant, a coordinated swarm.
The role:
We build and ship agent platforms that do real work in production. This role is for an engineer who designs self-running agent loops and multi-agent swarms, not someone who tweaks prompts one turn at a time. You will own systems that find work, do it, verify it, and report back, with humans at the gates rather than in every iteration.
This is a hands-on, ship-first engineering role. You write production code other engineers trust, and you treat agents as software systems that need triggers, budgets, stop conditions, verification, and observability like any other.
What we mean by loop engineering: designing an agent system as a control loop with an explicit trigger, scope, action, budget, stop condition, and report, plus a verifier that can reject bad work, so it runs unattended and knows when to stop or hand back to a human. Not a bigger prompt. A system that prompts the agents for you.
What we mean by swarm agents: a coordinated fleet of specialized agents (planner, builder, reviewer, verifier, scout, coordinator) with file and task ownership, shared state, quality gates, and clean handoffs, run by an orchestrator rather than one do-everything assistant.
What you'll do
- Design self-running loops. Define the trigger, scope, action, budget, stop condition, and reporting so an agent runs unattended, stays inside cost and iteration limits, and knows when it is done versus when to escalate.
- Build multi-agent swarms. Orchestrator plus specialized agents with clear file and task ownership, shared state or a shared mailbox, quality gates between stages, and handoffs that do not step on each other.
- Make verification first-class. Build the part of the system that can say no: the checks, evals, and reviewer agents that catch confident mistakes before they merge. A loop is only as trustworthy as its ability to check its own work.
- Own agent state and memory. Persistent on-disk state and per-turn context assembly so long-running tasks survive restarts and the system does not forget what the repo already knows.
- Ship the platform around the agents. APIs, services, queues, and integrations in TypeScript and Node, deployed to AWS, with real tests, tracing, and observability for long multi-iteration runs.
- Keep humans in the loop where it counts. Plan approval and pull request review, and active management of comprehension debt so the team understands what the swarm ships, not just that it shipped.
What we're looking for
- Strong engineering fundamentals. 5+ years writing production software that other engineers depend on. (Adjustable; we care more about what you have shipped than the number.)
- Hands-on loop engineering. You have designed agent loops with explicit stop conditions, budgets, retries, and self-verification. You can explain the difference between a task on repeat and a real loop, and you know why the verifier matters as much as the maker.
- Multi-agent or swarm experience. You have built or operated systems where multiple agents coordinate: orchestration, handoffs, shared state, ownership or locking, and quality gates.
- Fluency with modern agent tooling. Claude Code or Codex style agents, sub-agents, persistent memory and skills files, tool and function calling, MCP, and reason-act-observe loop patterns.
- Solid TypeScript and Node. Comfort with a service framework (NestJS or similar) and a typed data layer (Prisma or similar).
- Cloud and delivery. AWS (ECS or Fargate or similar), Docker, and CI/CD. You can take something from repo to production yourself.
- A verification mindset. You treat "done" as a claim to be proven, and you build the checks that prove it.
Nice to have
- Running 10+ parallel agents and managing token and cost budgets at scale
- Distributed systems, queues, and event-driven design
- React for agent-facing interfaces
- Prior work on developer tooling, orchestration frameworks, or internal agent platforms
- Familiarity with where loop engineering is heading next, including continual learning systems
- SOC 2 or ISO 27001 awareness for handling client data
How we work
- Plan, approve, execute. Agents propose, a human approves the plan, agents build, a senior reviews every pull request before merge.
- Small loops over big prompts. We design systems that discover work, do it, check it, and report, while we watch instead of type.
- Ship. Bias to production. Working software in front of users beats a perfect design doc.
- Direct and concise. Strong opinions, loosely held. Say the hard thing early.
What success looks like in 90 days
- You have shipped at least one self-running loop into production with a verifier the team trusts.
- You have extended or hardened a multi-agent workflow with clear ownership and quality gates.
- You have improved observability or cost control on an existing agent system so we can debug and trust long runs.
Tech environment
TypeScript, NestJS, React, Turborepo monorepo, Prisma, AWS ECS Fargate, Docker, GitHub Actions. Agent layer built on Claude and Claude Code patterns, MCP, sub-agents, and persistent memory.
Application Process:
To be considered for this role these steps need to be followed:
- Fill in the application form
- Record a video showcasing your skill sets
