# AI Teammate Infrastructure

> The infrastructure AI teammates need beyond an LLM: shared company context, identity, connected tools, execution, approvals, governance, audit, and durable operation.

**AI teammate infrastructure** is the layer around an AI model that lets an agent participate in real company work over time. The model supplies reasoning and generation. The infrastructure supplies the company-specific state, identity, permissions, tools, execution, and human control that make the agent usable as a teammate.

A chatbot can answer a question with a prompt and a model call. A teammate has a harder job: it must know what is changing, remember what matters, act through company systems, respect the identity of the person it is helping, and leave a trace that people can inspect.

## The stack behind an AI teammate

A practical stack usually needs these layers:

| Layer | What it answers |
|---|---|
| Model / harness | How does the agent reason and use tools? |
| Company context | What does the organization currently know? |
| Identity and permissions | Who is this agent acting for, and what may it access? |
| Connectors and feeds | What is changing in Slack, Jira, Gmail, databases, and other systems? |
| Memory | What should remain durable across sessions and agents? |
| Execution | Where can the agent run code or complete longer work safely? |
| Approvals and policies | Which actions require a person or a guardrail? |
| Audit and outcomes | What happened, why, and what was the result? |
| Delivery | Where does the teammate coordinate with people: Slack, web, email, MCP, or another surface? |

You can assemble those pieces independently. Lobu exists to make them one shared operating layer while still letting teams choose their own models and agent harnesses.

## Why the model is only one part of the system

Modern models are increasingly interchangeable at the application boundary. The difficult company-specific work is often elsewhere:

- connecting many internal systems once instead of once per agent
- keeping context current as those systems change
- separating one employee's access from another's
- giving agents durable state without leaking private scratch work
- brokering credentials so agent code does not hold raw tokens
- deciding which actions can run automatically and which require approval
- letting multiple agents hand work off without losing the organizational context

Those concerns remain even if you switch from one model provider or harness to another.

## AI teammate infrastructure is not the same as an agent framework

An agent framework usually helps you build the reasoning loop: prompts, tools, model calls, state machines, subagents, and retries. That is useful, but it is not the entire operating environment.

Lobu is designed to sit around existing agents as well as run managed ones. You can connect ChatGPT, Claude, Codex, OpenClaw, or another MCP client to the same company context, or build a proactive agent that runs under the same identity and approval model.

See [Lobu MCP](/mcp/) and the [architecture overview](/getting-started/architecture/).

## It is also not just an MCP server

MCP standardizes how an agent can discover and call external capabilities. It does not by itself decide:

- how company data is continuously ingested
- how organizational entities are resolved
- what should become shared durable memory
- how user-level permissions flow through the system
- how approvals, policies, and audit work
- how a background agent survives between interactive conversations

Lobu exposes MCP because it is a useful compatibility boundary. The underlying value is the shared state and governance behind that endpoint.

## The “company brain” is live, not a one-time index

An AI teammate needs more than documents embedded into a search index. Company state changes constantly: tickets move, people make decisions in Slack, a CRM field changes, an incident opens, a customer replies, a deployment completes.

Infrastructure for long-running teammates therefore needs an event path as well as a query path. Lobu connectors expose collected, streaming, or virtual feeds; durable events can update shared entities and activate behaviors. The result is company context that can move even when no one is actively chatting with an agent.

For the data model, see [Company context for AI agents](/company-context-for-ai-agents/) and [Shared memory for AI agents](/shared-memory-for-ai-agents/).

## Human control is part of the architecture

Giving an agent tools is not the same as giving it unrestricted authority. Production teammates need explicit control boundaries:

- scoped identities
- per-tool policy
- credential isolation
- human approvals for sensitive actions
- guardrails around input, output, tools, or egress
- append-only evidence and audit history

That control layer matters most when agents become proactive and operate outside a single supervised chat turn. See [Agent control plane](/agent-control-plane/), [guardrails](/guides/guardrails/), and [tool policy](/guides/tool-policy/).

## Where Lobu fits

Lobu is an open-source context and control plane for organizational agents. It combines:

- connected company systems and live feeds
- shared event-sourced memory and typed entities
- scoped user and agent identities
- MCP access for existing agents
- isolated execution for managed agents
- approvals, policies, credential brokering, and audit
- optional self-hosted or managed deployment

The agent itself is not locked to one model provider. The same company layer can serve several agent surfaces at once.

## When you probably do not need this layer

Use simpler infrastructure when you are building:

- one stateless chatbot
- one personal coding agent whose local files are the only required memory
- a deterministic workflow with no agent reasoning or organizational context
- one narrow MCP tool with no shared state, permissions, or long-running behavior

The extra operating layer becomes valuable when agents cross users, systems, sessions, and authority boundaries.

## Read next

- [Architecture and concepts](/getting-started/architecture/)
- [Shared memory for AI agents](/shared-memory-for-ai-agents/)
- [Company context for AI agents](/company-context-for-ai-agents/)
- [Agent control plane](/agent-control-plane/)
- [Related systems and research](/getting-started/related-work/)
- [Compare Lobu](/compare/)