AI Offices

160+ Agents in Production


A shared AI platform whose agents finish real work, not a chat window.

AI Offices

Agents that act on real operations, not just answer questions.

AI Offices is Buckler's multi-tenant agent platform, and I lead its engineering. It runs 160+ cloud AI agents in production, organised into 17 squads across 14 organizations, and serves Weblox, IT Ops and Work Balance as a shared AI service: a product asks for a capability, and the platform finds the right squad, runs the specialist on its model and bills the call.

What it changes for the business

How it is built

Cost and quality, measured

Platform

AI Offices runs on AWS EC2 with Linux, as Docker images run in rootless Podman containers with read-only file systems, behind NGINX with TLS. Agent media and signed deployment packages live in AWS S3, and avatars are served by CloudFront. On Google Cloud, Gmail changes reach the platform in real time through Pub/Sub push into AWS, and Google Chat and Workspace connect through least-privilege OAuth, all managed as code with Terraform.

AWS Google Cloud Linux

Language & Technology

The agent runtime is Python, FastAPI and SQLAlchemy on the AWS Strands Agents SDK, with tool calling, structured outputs, an MCP server and client, and A2A. The portal and admin apps are Next.js 16, React 19 and TypeScript, sharing the component, auth and state packages used across every Buckler product.

Python / FastAPI React / Next.js

Database

PostgreSQL on Amazon RDS with pgvector for RAG, so each client has its own knowledge base. Tenant isolation is enforced with row-level security, and migrations are managed with Flyway.


AI Integration

Agents are not tied to a single vendor: direct API access to Anthropic, OpenAI, DeepSeek and MiniMax, with routing per agent, cache and failover. Voice, video and web search are added through specialised APIs, and a person approves anything an agent publishes or sends.

Anthropic OpenAI DeepSeek MiniMax

CI/CD & Infrastructure

Shipped by IT Ops: the full test suite runs (pytest and Vitest), images are built and signed, and the host verifies the signature, runs the migrations and switches the containers. Credentials are encrypted with AWS KMS and secrets live in Secrets Manager. Infrastructure is code, with Terraform and Ansible.

IT Ops

Integrations

People sign in with Microsoft Entra ID or Google, with a TOTP second factor; agents use scoped JWT service tokens. Agents act on the systems a client already runs rather than on a copy of their data: Microsoft 365 through Microsoft Graph and Google Workspace, with OAuth so each person connects their own mailbox; Telegram, Microsoft Teams, Google Chat and email as channels; the office embedded in Weblox; and approved MCP servers and A2A agents.

Microsoft 365 Google Workspace

Key Features

Squads: a lead that delegates to specialists, each with its own tools.
Human in the loop: nothing is published or sent without a person.
Cost per conversation: every token billed to the organization that used it.
Multi-tenant: isolation at four layers, with one access check on every route.
RAG per client: knowledge bases that never mix between organizations.

Inside AI Offices


1 / 6
The Agent Office
2 / 6
Delegation and human approval
3 / 6
Agent management
4 / 6
Knowledge bases for RAG
5 / 6
AI benefits next to AI cost
6 / 6
AI Offices on AWS

The Agent Office — a squad (a Chief of Staff leading four specialists) as a live pixel-art office, with recent conversations, their cost and the delegated calls.
Delegation and human approval — the lead agent consults the finance and sales specialists, writes the weekly summary and waits for a person to approve before the email is sent.
Agent management — the lead agent's squad, who it delegates to, its capabilities and its configuration.
Knowledge bases for RAG — each organization has its own knowledge bases, with indexed documents, chunk counts and the embedding model.
AI benefits next to AI cost — conversations handled, share run with no person present, cost per conversation, response time and approval rate for sensitive actions.
AI Offices on AWS — how a request flows, where the data lives and how it is deployed.

Screenshots from a demo environment with invented data. Account identifiers, hostnames and client data are not shown.