Yugen Labs
What we build The problem Process Contact Book a call

From First Principles To Autonomous Systems.

We build agentic products, workflows, and internal systems for ambitious teams. We also build and operate products of our own.

Book a call See how we work
Yugen / Runner · Run · YG-0427
{{ headerStatus }}
6 Steps · 3 Tools · 1 Human Path
010203 040506 REQUESTVALIDATERETRIEVE REASONCOMMIT REVIEW retry 1/3dead-letter
{{ detail.request.text }}
{{ detail.validate.text }}
{{ detail.retrieve.text }}
{{ row.label }}{{ row.sub }}
{{ detail.reason.text }}
{{ row.label }}{{ row.status }}
{{ detail.review.text }}
{{ detail.commit.text }}
6 steps · 3 tools
1 human · 0 failures
1.42s total
{{ row.t }}
{{ row.text }}
{{ row.sub }}
What we build

What we build.

Three primitives. The same stack we run our own products on.
01 / n8n · make
triggervalidatellm stepcommit retry ×3page human idempotency keys · dead-letter queue

Workflow automation

Automations built on n8n and Make with LLM steps. Retries, idempotency, and a failure path that pages a human instead of failing quietly.

02 / typed tools
requestagenttyped ✓permit ✓ crm.readdb.raw ✕ typed args · allowlist only

Agents

Agents scoped to typed, permissioned tools. They do exactly what you allowed and nothing else. Not a chatbot with database access.

03 / auditable
queryembedindextop-kanswer source.pdf §14 answer → chunk_id → source

Retrieval systems

Retrieval you can audit. Every answer traces back to the source chunk it came from, so you can check it instead of trusting it.

We build these for clients. We also run two of our own on the same stack, and we are the ones on call when they break.

The problem

AI projects don't fail at the model.

Most of them die in the same four places.

01

The demo worked. Production doesn't.

Clean inputs in the demo. Real inputs in production. The gap between them is where the project stops.

02

The agent finished eight steps out of ten.

Nobody can tell which two failed, or why. There is no log to read.

03

A vendor changed a webhook payload.

The automation broke quietly. Someone noticed a week later, after the data was already wrong.

04

The build shipped and the person who understood it left.

Nobody on the team will touch it now, so it rots.

None of these are model problems. They are engineering problems.

Process

Four steps. No surprises in the middle.

Week 0 → Handover +30d
01 · CALL

Thirty minutes.

You walk us through the workflow that costs you the most time. We tell you whether a system can fix it and whether it is worth the money. Sometimes the answer is no.

02 · SCOPE

A written scope, a fixed price, a date.

We map the workflow against your real tools and your real volume. If the scope changes, we tell you before we build, not after.

03 · BUILD

Nothing gets ripped out.

We build around the tools you already use. You see working software every week instead of status updates.

04 · HANDOVER

Then you run it yourself.

You get the code, the documentation, and a working session with your team. We stay on call for thirty days. After that you run it yourself or keep us on retainer.

Contact

Tell us what's slow.

Send us the workflow that costs your team the most time. We will tell you whether a system can fix it.

Yugen Labs — 2026
What we build The problem Process hello@yugenlabs.com