Run agents on real enterprise work, every day.
You have built agent systems that people rely on in production. At TTC, own the runtime behind Runstate, where agents do paid work for enterprise clients every day, and set the engineering bar for the team that joins after you.
The Thinking Company works with CEOs, boards and private equity operating partners when AI needs to change the economics of a business. Our teams combine Transformation Partners, Forward Deployed Engineers (FDEs) and AI agents, working through Runstate, our internal delivery system.
Location Warsaw-first; Poland-based hybrid considered
Runstate already has users and paying work behind it. TTC's Transformation Partners and FDEs run client engagements through it, so an agent you ship meets real enterprise work within days, and you see where it fails.
The bet behind TTC is that agents can do most of the investigation, process design and build in a transformation, with senior people answering for the result. Whether that works at scale depends on the runtime you will own.
Your scope can grow from owning the runtime to leading engineering on Runstate and hiring the engineers who join after you. Sustained reliability, judgment and technical leadership can earn broader authority across the firm.
What makes it hard Agents at TTC work on one engagement for weeks, across many sessions, people and tools. That creates the engineering problems this role owns:
Long-running work must survive crashes, deploys, and waits for a person that can last days.
State must stay correct. A decision made in week two cannot quietly return in week five because an agent found an older document.
Most outputs have no single right answer. Judging a process redesign or a research finding takes evidence checks and regression sets built from real engagements.
Model choice and cost span several providers and are measured per run, against client budgets.
Agents work across many clients' material, and nothing may leak from one client to another.
The team is small, so coding agents do much of the engineering, and the harness around them sets the quality.
What exists today Runstate runs in production on a stack you would own and shape:
TypeScript (NestJS, Next.js) on Postgres, with Python for agents.
Temporal for durable background work.
LangGraph and LangChain Deep Agents, on LangSmith.
Several model providers behind one gateway, with EU-only processing for client data.
Tracing in LangSmith and Langfuse.
Claude Code and Codex as everyday coding agents.
What you will own
Make agent work durable.
Orchestrate long-running agent work so it survives failures, restarts and deploys, at the concurrency real engagements need.
Keep engagement state correct.
Every agent session, in any tool, starts from the current decisions, tasks and evidence, and changes them only through approved paths.
Run the model layer.
Route work across models and providers, set budgets, and know the cost and quality of every agent run.
Prove behaviour before release.
Build the tracing, evaluations and regression sets that show a change works before it reaches a client engagement.
Set the rules for human control.
With Transformation Partners and FDEs, decide where agents act alone and where a person must approve, then enforce it in the system.
Build the engineering loop.
Make coding agents part of how Runstate is built, with the tests, gates and review that keep their output safe.
Who you will work with You report to Bartek Pucek, TTC's founder, who works hands‑on on Runstate's product. You join a small engineering team, the Transformation Partners and FDEs who use Runstate every day, and the coding agents that do much of the build.
What the role requires Required
5+ years building and operating production backend or distributed systems.
At least a year running LLM or agent systems that real users depend on, including being on call for them.
Strong TypeScript and working Python.
Hands‑on experience with durable execution or workflow engines (Temporal or similar) and the failure modes of long-running jobs.
A practice for evaluating non-deterministic systems: tracing, eval sets and regression tests.
Daily use of coding agents, with a concrete method for verifying what they produce.
Strong signals
Production work with LangGraph, LangSmith or Langfuse, and routing across model providers.
MCP servers and tool design for agents.
Founder or first-engineer experience.
Systems that handle client, financial or health data under contractual or regulatory limits.
How you work You treat agents as distributed systems with non-deterministic parts, write decisions down, and want to build the firm as well as the code.
The team is small, with no platform, SRE or QA layer to absorb the edges. Client deadlines create intense periods, and production discipline still applies under time pressure.
Automatic disqualifiers
Your agent work stops at prototypes, and someone else ran it in production.
You build general frameworks before a real workload has proved the need.
You treat evaluation as manual spot checks.
You cannot explain the cost, failure rate or trade-offs of a system you built.
Limited frontend polish, research credentials and people‑management experience are acceptable. A strong distributed‑systems engineer with less agent experience can qualify through serious, current agent work they can show us.
Evidence we want to see We care about systems you owned and the decisions behind them. Be ready to discuss:
An agent or LLM system you built and operated after release, with its failure modes and how you measured them.
A long-running process you made durable, and what broke before you did.
A case where an agent acted on stale or wrong state, and the permanent change that followed.
How you know the cost per run of a system you built, and what you did to lower it.
An evaluation or regression set that caught a problem before release.
A rule you set for when an agent must stop and ask a person, and why it sits where it does.
Your current coding‑agent workflow and the checks that keep generated code out of production until it is verified.
A platform or framework you chose not to build, and what you did instead.
We assess the examples above more heavily than employer names or framework keywords.
Your first month In the first weeks you take over the runtime: Temporal workflows, agent workers, deployments and on‑call. You make engagement state hold across sessions, people and tools, so closed decisions stay closed, and widen the sources agents learn from.
You will ship new features daily.
Compensation and progression The role includes a path to leading engineering as the team grows, with broader technical authority and economic participation.
How we hire We reply to every application within 48 hours and aim to decide within two weeks.
A 30-minute call with Bartek Pucek, TTC's founder.
A 75-minute deep-dive on an agent system you built and ran: its failures, cost and evaluation.
A 60min vibe and culture check call.
A paid half‑day work sample on a small agent with a durability and state problem, using your own coding agents.
References from people who ran or extended a system after you handed it over.
A 60‑minute call with Bartek Pucek, TTC's founder.
Location and application Warsaw-first. Poland-based hybrid is possible for an exceptional candidate. In‑person teamwork is part of the role, with occasional time alongside client teams in Europe. Professional English is required.