Platform
Everything the installation does. With the real status of each feature.
Every feature below is either live on our own installation or marked as roadmap. We sell what the client receives on their VM — not a SaaS of ours.
Architecture
Where everything lives.
Teams is the front door. Everything else runs on your infrastructure, one container per agent, with the factory identity kept host-only.
Group 1 of 5
Coworkers, not bots
Each agent is a Teams app installed for everyone, with a name, personality and avatar. It answers DMs, takes part in channels, receives @mentions.
- Live
Named agents with personality and avatar, installed as Teams apps for every user; DMs and channels; @mention in channels.
- Live
Persona theme per company: naming convention and avatar style chosen at setup; the Chief of Staff proposes names and generates avatars within the theme, with approve / regenerate in chat.
- Live
Each agent has its own workspace (charter, memory, tools, skills, schedule, inbox, transcripts), versioned in git and backed up nightly.
- Roadmap
Off-site backup.
- Live
Agents remember: cross-session memory files, searchable transcripts of every conversation, and conventions written into their charter on admin request.
- Live
Attachments in; images, PDF and tables out (native Teams tables, Adaptive Cards with buttons).
- Live
Reminders and scheduled tasks in natural language (“remind me tomorrow at 9”, weekly briefs).
- Live
Agents message each other's humans: send a DM or a card to a colleague, with attribution and full context on reply.
- Roadmap
Voice (Teams calls) and email channels.
Group 2 of 5
Hire and manage by chat (the factory)
The Chief of Staff is the agent that hires the others. It interviews the manager, drafts the charter, generates the avatar and sends the Approve / Reject card. Nothing happens without the tap.
- Live
The Chief of Staff interviews the manager, drafts the new agent's charter, generates the avatar (approve / regenerate in chat) and sends an Approve / Reject card; on tap, registration, catalog publishing, container start and announcement are automatic.
- Live
Retire, grant or revoke DM access, promote or demote admins, set the announcement channel, link 1Password secrets to an agent (the value never passes through the model or the chat), change models — all in chat.
- Live
/help, /status and /new for every user; per-agent admin lists; domain-wide user access (*@company.pt).
Group 3 of 5
Governance and trust
The human gate is not a sales promise: it is a platform rule. Agents propose; people approve. And everything is logged.
- Live
Human gate on every consequential action (hire, retire, access, cross-channel sends): agents propose, humans tap.
- Live
Agents hold no cloud credentials; the factory identity lives host-only.
- Live
Secrets in the company's 1Password: a service account, kept host-only, pulls each secret and injects it into the container of the agent that uses it. Agents never hold that account's credentials nor see the value; they use it through the approved integration.
- Roadmap
Other secret managers besides 1Password (Bitwarden, Azure Key Vault).
- Live
Per-agent RBAC (admins / users / channels); channel access approved by an admin mention.
- Live
Container per agent (CPU / memory limits), read-only platform config, no access to other agents' folders.
- Live
Audit log of every command, grant and approval; usage log per run (tokens, cost, duration, model actually used); daily cloud-cost tracking.
- Live
Injection defence: a positional trust model — nothing a person writes can be read by the agent as a platform instruction.
- Roadmap
Second layer against injection: a decision model on your network classifies every untrusted content (attachments, emails, pages, tool results) before the agent reads it, and every proposed action before the approval card. It lowers the risk; human approval remains the control.
- Live
Verifiable ground truth: agents read a machine-generated platform-state.json (deployed version, recognised schedules, models, seats) instead of guessing.
- Live
Model management: per-agent model pin, effective next message; monthly deprecation watchdog.
- Roadmap
Kill switch per agent (< 1 min), emergency pause, working hours.
- Roadmap
Hash-chained audit export; Entra Agent ID registration; SSO for the admin portal.
- Roadmap
EU AI Act Art. 50 transparency banner (“You are talking to an AI coworker”) in every agent's app description.
- Live
Governance operating rhythm, built in: the Chief of Staff's daily brief (activity, costs, refused access attempts, pending decisions), the monthly review with us, and the audit / usage logs together form a documented, repeatable oversight process the client can show an auditor.
Daily brief — Tuesday
- 52 tasks handled by 4 agents; yesterday's cost €3.40.
- 1 access attempt refused: accounts@… asked to DM Quote Builder (not on the list).
- Seat 2 at 91% — I propose adding a seat next week.
- 1 pending decision ↓
- Requested by
- Sofia (accounts), yesterday 17:42
- Scope
- DM only; no channels
Group 4 of 5
Sovereignty and cost
Runs on your servers. Only Teams traffic crosses a zero-trust tunnel. The cost is fixed, per seat, and the client owns everything.
- Live
Runs on the client's on-prem VM (Ubuntu) or private cloud; only Teams traffic crosses a zero-trust tunnel.
- Live
Fixed monthly seats instead of per-token billing; seat usage reviewed daily by the Chief of Staff, who proposes when to add a seat.
- Live
Client owns everything: repo, containers, charters. No lock-in; hand-over documented (deploy runbook).
- Roadmap
Model choice incl. Azure OpenAI EU regions and open-weight models on a machine on your IT network (Mac Studio, DGX Spark, GPU server).
- Roadmap
Decision models on your IT network: they classify, sort and route without writing text; they pick the approval tier of each proposed action, never the approval.
- Roadmap
Per-message model selection: commands with no model, decisions on a decision model, drafts on a fast model, reasoning on the frontier model — within the ceiling the admin set.
Group 5 of 5
Agents that extend themselves
When an integration is missing, the agent builds it, documents it and we review it. A new platform capability reaches every agent at once.
- Live
Agents build their own tools and MCP servers, install packages persistently, and document what they can do.
- Live
A platform manual injected into every agent every message — new capabilities reach all agents at once; no retraining per agent.
Comparisons
Why not Copilot Studio, SaaS agents or Buzz?
Each alternative is good at what it was built for. We explain, without exaggeration, when each one makes sense.
vs Copilot Studio
Individual productivity inside M365, Microsoft models only, credits. We coexist and can register in Entra Agent ID.
Read the comparison →vs SaaS “AI employee” platforms
Minutes to start, huge catalogs — and your data, credentials and credit bill in a US SaaS.
Read the comparison →vs Block Buzz
The same agent-identity model — but in a new app, pre-1.0, with approval gates unfinished.
Read the comparison →