Metaminds
EN RO
Connect with us

Metaminds ResearchAI First · Internal agents

We put agents
on the team.

To keep pace in the AI era, we build and run our own agents — an ops copilot, a shared knowledge base, a bid assistant — each scoped to one job, grounded in our own knowledge, and safe by design. They share one reusable core, so the next one is weeks, not months.

3Agents already
in daily use
1Reusable core
they share
AI FirstHow the company
operates

01The idea

Keeping pace means
building for ourselves first.

The fastest way to understand where AI genuinely helps — and where it doesn’t — is to run it on our own work. So we do. Our engineers, our SREs and our bid team each have an agent built for their job, and we learn from them in production every day.

The point is not novelty. It is leverage: the same discipline we bring to a client engagement, turned inward, so the team ships faster and answers with its own knowledge instead of guesswork.

P1 We use what we build We don’t only deliver AI to clients — we run our own agents every day, so the team keeps pace with the tools it recommends.
P2 Grounded, not guessing Every agent answers from a version-controlled internal knowledge base, and cites where each answer came from.
P3 Scoped, not chatbots Each agent has one job and politely declines everything outside it. No open-ended, do-anything assistants.
P4 Safe by design Read-only where it matters, a human on the decisions that count, secrets redacted, sensitive work on models we host ourselves.

02The roster

Three agents,
three jobs done well.

dobby AI SRE · Ops copilot

Answers infrastructure and operations questions right in Slack — read-only investigation across our own systems, the weekly on-call rotation, CI and repository lookups, and knowledge-base answers. Scoped to ops, and never a general chatbot.

dobbypedia Internal knowledge base

The version-controlled brain the agents share: runbooks, patterns and organizational facts, curated by the team and read by every agent so answers stay consistent and current — treated as data the agents read, never as commands.

claudia Bid assistant

Reads an incoming tender and reports how we cover it — the right experts, comparable delivery experience and financial standing — mapped from our own record, with every claim cited back to its source, in a cowork-style workspace for the bid team.

03The pattern

One core, many agents

Every agent is a thin, well-scoped layer over the same foundation — a reusable conversation core, a registry of typed tools, and a security spine they all pass through. Build the base once; stand up the next agent by re-pointing it at a new job.

01
Agents one job each
dobbyAI SRE

Operations and infrastructure copilot, in Slack.

claudiaBid

Tender coverage analysis for the bid team.

uziDark factory

Ships software from an issue to a reviewed PR.

…and moresoon

New agents spin up on the same core.

02
Shared core built once, reused
Agent core

One reusable conversation loop, prompt and thread state every agent drives.

MCP tool registry

Typed, permissioned tools an agent is granted — nothing it is not.

Security spine

Secret redaction and allow / deny lists that every tool call passes through.

03
Grounding & models chosen per sensitivity
Knowledge base

Curated internal facts and runbooks the agents read from.

Self-hosted models

Open-weight models on our own GPUs for the sensitive work.

Frontier APIs

Best-in-class models for the data classes they are permitted.

04
Guardrails the same for all
Domain scope

Answers inside its job; declines the rest.

Read-only defaults

Look, don’t change — writes are the explicit exception.

Human-in-the-loop

People approve the actions that matter.

Auditable

Every action attributable and reviewable.

04Safe by design

Autonomy with the brakes wired in

An agent is only as trustworthy as its limits. Ours are built into the shared core, so every agent inherits the same guardrails — not a policy we hope holds, but a boundary the tools enforce.

Read-only by defaultInvestigative tools look, they don’t change. Writing anything is the exception — explicit, scoped, and attributable to a low-privilege bot.
Human-in-the-loopAgents propose; people decide. The actions that carry weight wait for a person to approve them.
Domain-scopedEach agent answers only within its job and declines the rest — no open-ended access, no do-anything assistant.
Secrets stay secretA redaction layer strips sensitive values before they can surface, and the most sensitive work runs on models we host ourselves.
Our boundary, our controlConfidential material stays inside our own systems. The agents are built for our team, on our infrastructure.

05What’s next

The first three,
of many.

dobby, dobbypedia and claudia are where we started, not where we stop. Because they share a core, the next agent is a matter of scope and knowledge, not a rebuild — a support triager, a delivery-docs assistant, a research scout, each a short step from the last.

One of them has already grown into a product of its own: uzi, our open-source AI dark factory, which takes a labelled issue to a reviewed pull request on its own. This is what AI First looks like from the inside — not a slogan, but a roster that keeps growing.

AI First, together

Build the next one with us.

We’re an engineering company that runs on its own agents — and we’re hiring the people who build them, and partnering with the teams who want the same. Let’s talk.