SERVICE — AGENTIC AI
Custom AI agents — built, run and extended by the engineer who wrote them
An AI agent (agentic AI) is software that takes a job, works through its steps with your tools and asks a human when it is unsure. We build yours on LangGraph, host it on EU infrastructure we operate, and extend it as your team's feedback comes in.
INa job with steps, tools and judgment
→
OUTan agent that runs it — supervised
The problem
Off-the-shelf agent platforms promise autonomy and deliver a demo. The job that actually needs doing — check the order in the ERP, look up the contract, draft the reply, book the slot, ask the sales lead whether the discount is fine — spans four systems and needs judgment at two points. That is a custom agent. And it needs someone who owns it after launch, not an agency that hands over a repository.
How an agent works
A trigger starts a run. The agent plans, calls your tools, checks its own result and stops at the human gate before anything irreversible happens.
FIG. 2 — AGENT LOOP: TRIGGER → PLAN → TOOLS → CHECK → HUMAN GATE → ACTION
What we build
- LangGraph agent core — an explicit graph of steps, tools and decision points — inspectable, testable, not a prompt loop that hopes for the best.
- Tool connections — read and write access to your ERP, CRM, mailbox, calendar and documents via APIs or Model Context Protocol (MCP) servers — scoped to least privilege.
- Human approval gates — the agent pauses before money, commitments or customer-facing sends and waits for a one-click approval in Slack, Teams or mail.
- Evals and tracing — every run is traced (Langfuse, self-hosted); a test set built from your real cases runs before every change, so quality can only go up.
- EU hosting we operate — Docker on Hetzner (Germany) or your own servers: model gateway, database, tracing, backups, updates — one operator for the whole stack.
What a run looks like
Every run is a trace: each step, tool call and approval with its timing. This is what you and we read when something goes wrong — and what the test set is built from.
RUN · ORDER-STATUS-AGENTDEMO DATA
- 00:00.0TRIGGERMail from customer: “Where is order 48213? Needed by Friday.”
- 00:00.4PLANlook up order → check shipment → draft reply → ask approval if the promised date is at risk
- 00:00.9TOOLerp.get_order(48213) → status: shipped, 2 items, carrier DHL
- 00:01.6TOOLcarrier.track("JD0148…") → in transit, delivery expected Thu
- 00:02.3CHECKpromised date met · no discount or refund involved
- 00:02.5DRAFTreply drafted (DE, formal), 86 words, cites tracking link
- 00:02.6GATEapproval requested in Slack #support — pending → approved after 41 s
- 00:44.1ACTIONmail sent · ticket #5531 set to “answered”
8 steps · 2 tool calls · 1 approval · 3.1k tokens
How it works
01
Job definition
Half a day with the people who do the job today. We write down the steps, the systems, the exceptions and the points where a human must decide.
02
Prototype agent
Within 2–4 weeks the agent runs real cases in shadow mode. You see every step and tool call in the trace and grade the outcome.
03
Go-live with gates
Approval gates, permissions, alerting and the test set are locked in. The agent takes the clean cases; your team keeps the rest.
04
Operate & extend
Monthly: we read the traces and your feedback, fix what failed, build the next capability — and ship it against the test set. The agent grows with how you use it.
Build, run, extend
Three phases, one owner. The first is a project; the other two are a monthly engagement you can end any month.
Weeks 1–4 · project
BUILD
A working agent on your real cases, traced end to end.
- Job definition with the people who do it today
- LangGraph graph: steps, tools, approval gates
- Tool connectors (API / MCP) with least-privilege credentials
- Test set from real cases; shadow-mode runs you grade
Deliverable — a working agent on your cases, every run traced
Monthly · cancel any month
RUN
A system someone is responsible for.
- Hosting on Hetzner (Germany) or your servers, operated by us
- Monitoring, alerting, backups, model and dependency updates
- Review of failed and escalated runs
- Monthly cost and quality report
Deliverable — uptime, a named operator and a report
Monthly · same engagement
EXTEND
An agent that grows with how you use it.
- Your team's feedback becomes a prioritized list
- Next capability built, tested against the eval set, shipped
- Autonomy widened per action type where traces show it is safe
- Documentation kept current; your team can take over any time
Deliverable — a new capability every month, code and data yours
Specifications
| Typical stack | LangGraph (Python), Postgres + pgvector, Langfuse tracing, EU-hosted LLM (Mistral, Azure OpenAI EU) or local model via Ollama / vLLM |
|---|---|
| Hosting | Hetzner (Germany), your servers, or hybrid — operated by us |
| Integrates with | ERP, CRM, Microsoft 365 / Google Workspace, ticket systems, any REST API, MCP servers |
| Prototype timeline | 2–4 weeks; then monthly operate & extend |
Chatbot, workflow or agent?
Three different tools. We build all three and pick by the shape of the job, not by what is fashionable.
| Chatbot | Workflow | AI agent | |
|---|---|---|---|
| Gets | a question | a trigger | a goal |
| Steps | one answer | fixed sequence | chosen at runtime |
| Tools | reads documents | calls APIs in a fixed order | reads and writes — decides which |
| Exceptions | escalates to a person | fails over to a fallback path | re-plans or asks a human |
| Human gates | optional | at fixed steps | before risky actions, by policy |
| Best for | FAQ, support tier 1 | recurring, predictable processes | jobs with branches and judgment |
| Our service | AI customer support | Workflow automation | this page |
Frequently asked questions
What is an AI agent — and how is it different from a chatbot or a workflow?
A chatbot answers questions. A workflow runs a fixed sequence of steps. An agent gets a goal, decides which steps and tools to use, checks its own results and asks a human when unsure. Agents are right for jobs with branches and judgment; for a fixed sequence a workflow is cheaper, and we will say so.
Why LangGraph and not an agent platform?
LangGraph is open source, runs on our own infrastructure and makes every step explicit — so we can test, trace and change the agent without a vendor in the loop. Platforms are faster for a demo and slower the moment you need something they did not anticipate.
What does "operate and extend" mean in practice?
A monthly engagement: hosting and monitoring, model and dependency updates, a review of failed or escalated runs, and a prioritized list of improvements from your team's feedback — built, tested against the eval set and shipped. You can stop any month; code and data are yours.
Can the agent act on its own?
Only where you decide. The default is supervised: reads are free, writes and sends wait for approval. Autonomy is widened per action type once the traces show it is safe — and narrowed again with one setting.
What does an AI agent cost?
Two parts. The build is a fixed price agreed after the free discovery call, sized to the job, the systems involved and the number of approval points. Operate & extend is a monthly fee sized to the hosting footprint and the pace of change you want — cancellable monthly. Model usage is billed at cost. We publish no list prices because no two agents are the same scope.
LangGraph or n8n — which one do we need?
n8n when the steps are fixed and visual workflows are enough; your team can maintain them. LangGraph when the agent has to choose steps and tools at runtime, keep state across turns and pause at approval gates. Many systems use both: n8n for triggers and integrations, a LangGraph agent for the judgment in the middle.
CONTACT — HELLO@OLTEN.AI
Get a free 30-minute AI audit
We look at your processes, name the use case with the fastest payback and tell you honestly if AI is not the right tool. No slides, no obligation.
Or write directly: hello@olten.ai