AI agent journey
Design your AI agent in nine questions
We start every project in your shoes, not in a slide deck. Answer the questions a forward deployed engineer asks in the first meeting, and see the agent, the model and the hosting we would build for you.
Before you start
How an AI agent actually works
A chatbot answers a question. An agent finishes a job. It reads your data, decides what to do, uses your tools to do it, and checks the result before anyone relies on it. Each turn of that loop is logged.
Read
The agent collects what it needs from the systems you allowed: the inbox, the ledger, the ticket queue, the repository. It sees only what its permissions allow.
Reason
A large language model works out what the job needs and plans the steps. This is the only place intelligence is used, so it is spent where judgement matters.
Act
The agent uses real tools through connectors and MCP servers: it drafts the reply, files the report, raises the pull request. Rule-based steps stay as plain code.
Check
Before anything is sent, spent or deleted the agent stops at an approval gate. You approve, or reject with a reason it learns from. Every step is traced.
Three layers in every agent we build
Your knowledge
The data sources we connect during discovery, scoped to the same permissions your staff already have.
The model
Claude, GPT, Gemini, Grok or an open-weight model, chosen for the task, the data and where it must run.
Control
Approval gates, budgets, stop conditions and an audit trail, run on the Jobshout agent framework.
Discovery
We get into your shoes first
Most AI projects fail because they automate the wrong step. Our forward deployed engineers sit with the people who do the work before writing a line of code.
Day 1
Shadow the work
We watch a real working day: the inbox, the approvals, the spreadsheets that hold the business together. We note every hand-off and every wait.
Day 2
Map the data sources
For each step we record where the data lives, who owns it, how it is reached (API, export, database, email) and how sensitive it is.
Day 3
Find where judgement is needed
We split each process into rule steps, which stay as ordinary automation, and judgement steps, where an LLM earns its cost. Agents go only where intelligence is required.
Week 1–2
Build and run beside you
The first agent runs on your real data with an approval gate on every risky action. We tune it with the people who use it, then widen its permissions only when they trust it.
The journey
Tell us about your business
These are the questions we ask in the first customer meeting. Your answers stay in your browser until you choose to send them.
Real-world scenarios
What we hear in customer meetings, and what we build
Composite examples from the kind of first meetings our forward deployed engineers run. Names and numbers are illustrative.
CEO · professional services
"I spend my first two hours every day on email."
The FDE mapped Outlook, the CRM and the board portal, and found that most replies followed five patterns. An Executive PA agent now drafts replies and meeting briefs overnight; the CEO approves from a phone.
- Data: Microsoft 365, HubSpot, board portal
- Model: Claude through the company's Azure tenant
- Control: every reply approved before it is sent
CFO · manufacturing
"Month-end takes nine days and most of it is copying."
Discovery showed the judgement was only in the variance commentary; matching and posting were rules. We automated the rules in code and put an agent on the commentary and the policy checks.
- Data: Sage, bank feeds, purchase order spreadsheets
- Model: a mix of a local model for extraction and a frontier model for commentary
- Control: no posting or payment without a named approver
CTO · software company
"My senior engineers spend their week reviewing other people's code."
A PR review agent checks every pull request against the team's standards and a second agent writes incident summaries from Sentry and logs. Seniors review the agent's notes, not the whole diff.
- Data: GitHub, Jira, Sentry, Slack
- Model: Claude Code for review, with an open-weight model for log triage
- Control: comments only; merging stays human
Law firm · offline
"Nothing privileged can ever leave the building."
We sized AI workstations with Workstation and ran open-weight models fully offline. A contracts agent compares documents to the firm's playbook and drafts redlines; partners approve each one.
- Data: document management system, precedent bank
- Model: open-weight models on local GPU workstations
- Control: air-gapped, with audit logs kept on site
Forward deployed engineers
An expert in your team, not a ticket in a queue
Every project gets a forward deployed engineer (FDE). They join your meetings, learn how your business really works and build and deploy the agents with you, fast.
Asks the right questions
Where does the data live, who owns it, what happens when it is wrong, and which step really needs judgement.
Builds in your systems
The agents live in your repository and your cloud or your hardware, connected to the tools your team already uses.
Stays until it works
They tune the agent with the people who use it and hand over runbooks, so your team can run it without us.
Certified across the platforms you might choose
- Claude Code (Anthropic)
- OpenAI
- Google Cloud and Gemini
- Microsoft Azure AI
- xAI Grok
Framework
Built on Jobshout, our agent framework and task manager
Jobshout is the open-source framework every TenthMatrix agent runs on. Each agent owns a lane, uses its own tools and stops at the same approval gate. Every run lands on a task board with its cost and trace.
- Task board with projects and sprints
- Multi-agent workflows as a dependency graph
- Scheduled jobs for daily and weekly work
- Skills, plugins and MCP connectors
- Bring your own model, including local models
- Per-run cost, latency and OpenTelemetry traces
Self-hosted AI
Want your AI with no internet connection at all?
Some data can never leave the building. We run open-weight models on AI workstations inside your network, fully offline, and build the same agents on top. The hardware comes from our sister company Workstation.
Questions we are often asked
What is the difference between a chatbot and an AI agent?
A chatbot answers. An agent completes a task: it reads your systems, plans, uses tools to do the work and stops for approval before anything risky. You get a finished draft, report or update, not a paragraph of advice.
Do we have to choose one AI company?
No. Many of our clients use a frontier model for the hardest steps and a small open-weight model for routine ones. Jobshout lets each agent use a different model, and you can switch later without rebuilding.
Can the agents run without any internet connection?
Yes. With open-weight models on AI workstations from Workstation, the whole system, including the models, the framework and the logs, runs offline inside your building.
What does a forward deployed engineer actually do?
They sit with your team, map your processes and data sources, decide where an LLM is worth using, and build and deploy the agents in your environment. They stay until the people who use the agent trust it.
What happens to the answers I give in this journey?
Nothing, until you press send. The blueprint is worked out in your browser. If you send it, it reaches our team as an enquiry so an engineer can prepare for the first meeting.
Prefer to talk it through?
Book a discovery call and a forward deployed engineer will run the same questions with you, live, on your own processes.