# Agents That Do the Work and Show Every Step.

A VectorMind agent takes a job, uses the tools you give it and logs what it did. You decide what it can reach and where a person signs off.

## How an Agent Connects

Work comes in, the agent uses only the tools you allow, and risky actions wait for a person.

Work arrives through channels, files and skills. The agent calls toolkits and MCP connectors, pauses at approvals for a person, and every step lands in the run log.

- **Channels**: Chat, email, WhatsApp and webhooks bring work in, and answers go back the same way.
- **Files**: Documents people attach in the chat, read as they are or with fields pulled out by Iris.
- **Toolkits**: Built-in tools you add per agent, from web search to a scheduler.
- **Approvals**: Any tool can need approval. The run pauses until a person approves or denies the call.
- **MCP Connectors**: Your organisation's own MCP servers, each listing exactly which tools the agent can call.
- **Credentials**: Kept in their own store and attached to connectors, never written into an agent's instructions.
- **Run Log**: Every step, tool call and result is kept, so you can open any run again later.
- **Skills**: Approved instructions your organisation writes once and reuses across agents.

## What an Agent Is

Instructions, a model and the tools it may use, saved as a versioned spec. You build it on a canvas and watch each step as it runs.

Agents are built on Praxis, the agent framework in VectorMind. /products/praxis

- **Chat**: Talk to it in VectorMind and follow each step in the side panel.
- **Email**: Any IMAP mailbox or Microsoft 365. It reads, and replies if you allow it.
- **WhatsApp**: It answers messages sent to a connected number.
- **Webhook**: Your own systems call it with a key an administrator issues.

## How It Stays Governed

The controls sit in the platform, not in the agent's instructions.

- **Tools That Change Things Start Off**: Tools are switched on per agent. Scheduling and a skill's own tools ship off, and toolkits that run code wait for a sandbox.
- **Nothing Goes Live Unapproved**: An administrator approves each new version, and the approved one keeps serving while a change waits.
- **Credentials Stay Locked Down**: A credential can be limited to named agents, and only an administrator can revoke one. An agent on an OAuth account gets only that account's permissions.
- **Roles Decide Who Does What**: Building and publishing need a SuperUser; approving versions, issuing webhook keys and revoking credentials need an administrator. New agents stay private until shared.
- **Retired, Never Erased**: A retired agent keeps its full history, and an administrator can restore it.

## What It Connects To

Each kind of reach has its own controls.

### Toolkits

Built-in tools you add per agent.

- Read files people upload
- Extract fields from documents with Iris
- Search the web
- Keep named memories, contacts and a to-do list
- Ask the person a question mid-run
- Schedule a follow-up, or wait and resume
- Hand work to another agent or to sub-agents

### MCP Connectors

Register a server in your organisation's catalogue and attach it with a credential. Requests for new servers go to an administrator.

### Skills

Each version is used only once an administrator approves it. Skills are text, so they can't run code.

## Agents by Industry

What teams build in the sectors we serve, using only the tools on this page.

### Contract Reviewer (Legal)

Compares each clause of an uploaded agreement with the firm's standard positions in an approved skill.

Uses: Files, Skills.

A lawyer decides what to push back on.

### SOP Assistant (Pharmaceutical)

Answers cold-chain and dispensing questions from approved SOP skills, naming the SOP behind each answer.

Uses: Chat, Skills.

QA approves each new SOP version first.

### Export Document Checker (Agriculture)

Pulls consignment details from certificates and packing lists with Iris and lists any mismatch.

Uses: Iris, Files, To-do list, Ask the user.

The export manager confirms before a consignment is marked ready.

### Inspection Planner (Mining)

Keeps a safety audit's to-do list, finds people on the contact list and schedules a follow-up with each.

Uses: To-do list, Contacts, Scheduler.

A person approves every follow-up.

### Inbox Triage (Financial services)

Sorts failed debit order queries from routine mail and answers routine questions from an approved FAQ skill.

Uses: Email, Skills.

Replies can stay off until the team is ready.

## See an Agent Run on Your Own Work.

Bring one job your team repeats, and we'll show you how an agent would handle it in VectorMind, step by step. Book a briefing: info@cohesionx.co.za