How to set up your multi-agent company on your laptop

A useful AI team starts with a file structure you can understand. Set up one company folder on your laptop, with governing briefs, clear roles and current-work records that help your AI chats work toward the same objective.

At Mevanu, company guidance, focused roles and retained work records help us coordinate our own work. Here is a small version you can set up yourself. It uses one Executive Partner and four helpers, with you carrying the relevant material between chats. No coding or automation is needed for this example.

Start with your business

Consider our fictional copywriter: ten years of experience, a job she needs to keep, and six hours a week to explore a business of her own. She begins with her expertise, the kinds of clients she understands and material she has permission to use. Employer information and confidential client work stay outside the example.

Her first note can be simple: what she wants to offer, whom she might help, the time available and what she has yet to decide. An uncertain customer group belongs under “to explore.” Giving an AI a clear uncertainty helps it ask a useful question instead of treating a guess as settled business direction.

Give your EP and helpers clear roles

An LLM, or large language model, is the technology behind many AI chat tools. For this setup, a helper is a chat with a focused assignment. More automated systems can also give AI assistants tools to act on their assignments; here, you manage the handoffs yourself.

Your Executive Partner, or EP, is the continuing AI assistant you ask to coordinate the work. It helps maintain the overall objective, prepares assignments, integrates results and brings consequential decisions back to you. You remain the business owner.

These are roles you can use within the features and usage limits of your existing account. They do not require five subscriptions. Call on a helper when its separate perspective or focused context is useful. Every venture has its own evolving needs; adapt these roles and files as your work develops. Anthropic’s guide to using multiple agents explains both the benefits of specialization and the coordination overhead it introduces.

Create one company folder

The root directory is simply the main folder containing this company’s work. Start with a structure you can understand at a glance:

Company/
├── README.md
├── AGENTS.md
├── COMPANY_BRIEF.md
├── EP/
│   ├── ONBOARDING.md
│   └── CURRENT.md
├── Researcher/
├── Analyst/
├── Writer/
├── Assurance/
├── references/
└── archive/

The .md ending means Markdown: a plain-text document with simple headings and lists. You can write it in a text editor. These names are suggested conventions; each file earns its place by doing a distinct job.

Save a source once and link to it where possible. Keep personal material unrelated to the business outside this company folder. Check the access you grant in your AI tool; a folder name such as “Assurance” does not set permissions.

Ask your first EP to prepare its onboarding

Tell your first EP what the business is, what you want to accomplish next and which existing material it may use. Ask it to draft the instructions it will follow. You can copy this starting prompt:

Help me set up a simple company workspace and act as my continuing
Executive Partner. I am the business owner.

Business purpose: [describe it]
People I may serve and my proposed offer: [include uncertainties]
Current objective and time available: [describe them]
Existing material you may use: [list it]
Decisions I keep, privacy limits and spending limits: [state them]

Inspect only the material I have authorized. If you cannot read my
local files, ask me for the relevant text or uploads.

Draft your own onboarding prompt and reading sequence. Propose a
minimal company folder, company brief, AGENTS.md instructions,
current-work record and role briefs for a researcher, analyst,
writer and independent assurance reviewer.

Explain each part. Preserve existing records, identify missing facts
and ask only questions that materially affect the setup. Use helpers
when useful, with clear assignments and manual handoffs where needed.

Show me the proposed setup before making changes. Once I authorize
it, complete the agreed steps within your actual access, identify
anything I must save myself, and verify the resulting files and links.

Read the proposed onboarding prompt and correct it. Check that it captures your intentions and limits. Then ask the EP to adopt the agreed instructions and help establish the workspace. You can continue with the same EP chat.

Where direct file access is unavailable, save the generated documents yourself. Paste or upload the relevant instructions into the helper chats you use. Creating folders does not make an ordinary web chat read them or move their contents between conversations.

In Codex, AGENTS.md has documented discovery rules based on the project and working directory. Other brief filenames need explicit reading instructions. Check the official AGENTS.md documentation when applying this setup in that tool.

Carry the work between chats

In this manual arrangement, the copywriter is the coordinator between conversations. EP can prepare a compact assignment containing the goal, relevant context, source material and expected result. She copies that assignment into the helper chat, then brings its result back to EP.

EP assignment → you → relevant helper
Helper result → you → EP
Draft + original brief + sources → you → independent assurance
Assurance findings → you → EP
EP’s revision assignment → you → writer, when needed
Corrected draft → you → assurance for recheck → you → EP

Keep a link or filename with each handoff so you know which version is being discussed. In a chat without access to that file, include the relevant contents too. After the review is resolved, save the agreed output and update the current record. You do not need every helper for every piece of work.

Make assurance independent of the drafting

Give the assurance reviewer the actual draft, the original request, the acceptance criteria and the supporting sources. Ask it to identify consequential errors, omissions, unsupported claims and departures from the brief, with evidence for each finding. Let it form its assessment before adding the writer’s explanation of why the draft is good.

For example, if a draft company brief gives our copywriter ten hours a week, assurance should point to her original six-hour limit. She carries that finding back to EP, which prepares a focused correction. The writer revises the draft, and assurance checks the corrected version. This is an illustrative check, not a reported customer result.

Keep the reviewer’s findings separate from the author’s revisions. EP resolves supported findings or brings an unresolved disagreement back to you. A confident reviewer, or agreement between two chats using the same model, can still be wrong; check the cited evidence. The independent role creates room to challenge the work before you accept it.

At Mevanu, independent challenge is part of how we review our work. For software-specific checks, our code review guide develops that discussion further.

Keep useful context and continuity

Stronger input = stronger output is a useful starting principle: give an assistant a clear objective, relevant sources, constraints and a way to judge the result. That gives it a better basis for useful work.

The information available to a model at a given moment is its context. Models process text in units called tokens. Repeated explanations, irrelevant documents and correction loops can use more context and effort. A focused helper brief can keep unrelated material out of a task; a retained decision can prevent the same question being reopened.

Anthropic’s context engineering guidance explains why selecting and maintaining relevant information matters. Clear inputs and retained decisions aim to reduce avoidable rework. Extra chats and handoffs also consume time and tokens, so judge the setup by the useful results and total effort it produces. Five roles are a starting example, not a promise of lower costs.

Before leaving a task, have EP propose a short update to the current record: the objective, decisions accepted, exact output and review references, what remains uncertain and the next action. Save it. On returning, supply that record and refresh anything that may have changed. This gives the next session a place to resume without rereading every old conversation.

Take the first step

Create your company folder and write a few sentences about its purpose, possible customers, current objective and limits. Give those to your first EP with the setup prompt.

Then ask it to read back its role, the sources it actually received and the objective it is pursuing. Check its answer against your saved brief. You now have a small company workspace you can understand, inspect and improve as the business develops.

If you would like help, a consultant can tailor the setup to your budget, available time, complexity and ambition. Together, you can draft a working roadmap: what to expect first, what comes later and when, with the assumptions and dependencies behind those expectations. Update it as the venture develops.

Have a question about a setup for your own work? Talk with Mevanu. Scope, deliverables and fees are agreed before work begins.