Skip to content
03AI Handbook

The single source of truth.

One place for how we use AI, who can build what, and where the line sits. It replaces the scattered docs — curated, current, and short on purpose.

01Day-to-day

Using AI day-to-day

Individual use is open to everyone. Pick the tool that fits the task and go — the only thing to watch is the data.

Approved toolsOpen to all
  • Claude

    Our default for writing, thinking, and long-document work.

    Assistant
  • Claude Code

    Agentic coding in the terminal — a heavily-used default for real engineering work.

    Coding
  • ChatGPT

    A strong second opinion and quick everyday drafting.

    Assistant
  • Perplexity

    Answers with sources when you need to cite where it came from.

    Research
  • Cursor

    AI-native editor for building and refactoring in your own repo.

    Coding
  • GitHub Copilot

    Inline completion inside the editor you already use.

    Coding
Something you’re using, or want to try?A first-draft list — propose a tool to explore or share with the team.

Yours to use.

Using these is open to everyone — no request, no approval. The only thing worth a second thought is the data: what’s in the prompt, and where it’s going. Keep client data, secrets, and anything you wouldn’t email outside the company out of third-party prompts, and build freely.

02Governance

Who can build what

This is the central distinction. Using AI in your own instance is open to everyone. Wiring an agent in — deep workflow integration, multi-agent systems, or standing access to a shared system — is a different class of decision.

That second class gets a closer look together — technical oversight and review. Trusting a colleague and granting their tools standing access are two different decisions, and the second often involves tradeoffs outside any one person’s expertise, so we make it as a team.

Decision flow

Can I wire this agent in?

Start · a new AI tool or agent
  1. 01

    Does it get standing access to a shared system?

    Email, calendar, deals, money, the CRM — anything beyond your own sandbox.

    No — it’s just me using AI
    Yes
  2. 02

    Can it act on its own — send, reply, change, or pay?

    Not just read or draft for you to approve, but take the action itself.

    No — read or draft only
    Yes
  3. 03

    Does it run unattended, or touch other people’s data?

    A standing integration, a multi-agent setup, or anyone’s inbox but your own.

    No
    Yes
Open

Go ahead

This is individual use. Mind the data, and build.

A closer look

Let’s look at it together

Deep integration, autonomous action, or other people’s systems. We look at the setup before it’s switched on — the tricky parts often sit outside any one person’s expertise.

Hermesco-workOpen-Claw-style VAs

Three questions. Any “no” to standing access, autonomous action, and shared data means go ahead. A “yes” just means we take a closer look together — at the integration, before it’s switched on.
03Agents as teammates

An agent gets its own identity, like any teammate

When a tool starts acting on its own, we set it up the way we’d onboard a person — its own identity, scoped to the job — rather than letting it run on our personal access. Here’s what that looks like in practice.

  • Its own identity

    An automated worker gets its own account and credentials, like a new hire — it never runs on someone’s personal login or inbox.

    A named identity we can see, manage, and switch off — not a person’s access borrowed in the background.

  • Only the access its job needs

    Scope follows the role. A drafting helper can draft; it doesn’t get the keys to billing because it happened to ask.

    Least privilege by default. Every permission is granted on purpose, not inherited from whoever set it up.

  • A person stays on the loop

    For anything that leaves the company or changes a record, a teammate reviews and approves. The agent proposes; a person decides.

    It assists and drafts. People stay accountable for what goes out under our name.

  • A clear, attributable trail

    Like any teammate, its actions are logged and attributable — we can always see what it did and why.

    Activity stays visible and reviewable. No silent, unattributable changes.

  • Onboarded, and offboarded

    Access is granted when the agent starts a job and removed when the job is done — not left running indefinitely by default.

    Standing access is reviewed and easy to revoke, like any account we manage.

04Pace by context

We set our own pace

The lab moves quickly — adopting and automating early is part of the job here. That pace is a choice we make in the lab, and it doesn’t automatically carry to every context we work in.

The same care applies everywhere. What changes is how far we lean in: wide open in the lab, and steadily more deliberate as the work gets closer to clients.

By context
  1. 1

    In the lab

    We move fast

    This is where we experiment and wire things in early. It’s internal and ours to break, so we lean in and learn quickly.

  2. 2

    Shared, day-to-day systems

    More deliberate

    The tools and systems the team relies on. We bring in what’s proven in the lab, a little more carefully.

  3. 3

    Anything client-facing

    Most careful

    Where our work meets clients and partners, we move slowly and deliberately. The bar is highest where the stakes are.

One pace doesn’t fit every context. We move fastest in the lab, so the patterns are proven by the time they reach client-facing work.
A living reference

Curated, owned, and growing.

This is a handbook we own and grow — not a wiki everyone edits. It stays short on purpose: one current page beats ten stale ones. When the work outgrows what’s written here, we add to it.