Running an open-source AI agent on your own machines
An agent that works without depending on a vendor. In one day, your technical teams install an open-source agent and a local model on their own machine, give it real work, write reusable procedures and set the guardrails. The question "what leaves this machine?" is asked at every step, and answered by checking.
Target audienceCIOs, CISOs, developers, system administrators, ops, AI leads: the technical people who equip the teams
What you'll be able to do
Tell a local agent from a hosted assistant, and name what leaves the machine in each case
Install an open-source agent and the local model suited to the hardware at hand
Give the agent a complete work task and check what it produces
Write a reusable procedure and decide when a separate agent profile is warranted
Set the guardrails for production use: approval, access separation, logging, pinned version
Programme
- Morning
What a local open-source agent is
- Assistant, agent, local agent: what each one does and does not do
- The learning loop: the agent builds its own procedures, and what that means for review
- What stays on the machine and what leaves: local model, tool gateway, web search, session summarisation
- Demonstration: the same task run by a local agent and by a hosted assistant, network traces compared
- Morning
Installation and choosing the model
- Installation on the participant’s own machine, desktop mode
- Reading your hardware: available memory, acceleration, a sensible model size
- Two models compared on the same work prompt: quality, speed, language
- The settings that matter: context length, temperature, where the working files live
- Morning
The first useful agent
- A real task from the participant’s own team
- Writing the prompt: context, source, expected format, success criterion
- A guided run, then the participant reviews what came out
- Graded exercise: a second task run solo, end to end
- Afternoon
Skills and profiles
- A reusable procedure becomes a skill: structure, invocation, updates
- When a separate agent profile is warranted: its own memory, its own access, its own authority
- A decision tree and an eight-question test, handed to each participant
- Exercise: sorting four team needs into prompt, skill, sub-agent or profile
- Afternoon
Automating
- Background tasks: starting them, tracking them, being notified without watching
- Driving the browser on an internal web application, and what to forbid it
- Surfaces: desktop, internal chat, command line
- Exercise: a recurring team task scheduled and triggered
- Afternoon
Guardrails and running it
- Approval before anything goes out: sending, publishing, writing into a third-party system
- Access separation: what the agent reads, what it must never see, where secrets live
- Logging: what is recorded, where, and for how long
- Release pace: why to pin a version in production, and how to run an upgrade
- AI Act register, and the team usage note, written during the session
Practical information
1 day (7 hours)
CIOs, CISOs, developers, system administrators, ops, AI leads: the technical people who equip the teams
Everyday use of a computer and a terminal. No development skills required. Hardware: Apple Silicon Mac or recent PC, 16 GB of RAM, 30 GB of free disk space, administrator access. Compatibility and installation permission are checked during the needs analysis.
1 to 10 people
€1,800 to €4,000 excl. VAT per group
OPCO funding possible.
Participants install on their own machine, never on a demo workstation. The exercise tasks come from their own team and are collected during the initial assessment. One thread runs through the day: at every stage, which data leaves the machine, to whom, and in what form.
Entry assessment, including a hardware check. One graded exercise per session, six in all. At the end, a before-and-after self-assessment of skills, then the trainer validates each learning objective (acquired, in progress, not acquired), recorded in the end-of-training certificate.
Colombani.ai, a Qualiopi-certified training provider. The trainer is certified by Anthropic, holds certifications issued by OpenAI, and deploys agents on client engagements every week.
At least 2 weeks between enrolment and the start of training.
Qualiopi-certified provider (certificate QUA009472), registered under training provider number 11757549975: OPCO or FIF-PL support can be requested. The decision rests with the funding body, based on your sector, its criteria and its budgets. Colombani.ai prepares the application with you.
Price per group, €1,800 to €4,000 excl. VAT, depending on group size and format: charged per group rather than per person. OPCO or FIF-PL funding can be requested depending on your status; the amount left to pay depends on the sum granted by the funding body.
Accessibility
This course is accessible to people with disabilities. Contact the disability officer in advance to discuss possible adjustments.
Ulysse Trin, [email protected]
Post-training support
Frequently asked questions
How does this differ from the "Local AI" course? +
"Local AI" answers "my data must not leave": the trainer installs everything, participants query their own documents, no terminal involved. This course answers "I want an agent that does the work": participants install it themselves, give it tasks and automate them. The audiences differ, and the two fit together within one organisation.
Is a paid subscription required? +
No. The agent and the models used are open source, the agent under the MIT licence. A subscription to a hosted model is optional and only serves the comparison made at the start of the day.
Our machines are managed by IT. Is that a blocker? +
That is settled before the session, not on the day. The needs analysis checks the hardware and the permission to install. Failing that, a loan machine or a lighter configuration is offered.
The tool moves fast. Will the materials be current? +
The version used is pinned for the session and recorded in the report, and the materials match that version. The day teaches practices that outlive releases: scoping a prompt, checking a result, setting a guardrail.
Related courses
Local AI training: your data stays with you
Install and use AI on your workstation, without sending data to the cloud. Ollama, AnythingLLM, RAG on your documents. Digital sovereignty starts here.
AI agent security training
The real risks of autonomous agents: prompt injection, exfiltration via MCP, sandboxing, permissions, security hooks. AI agent red teaming.
Got a project in mind?
Describe your situation. The first 30-minute call is free. Get a frank response within 48 hours.