Creating Agents

Create an agent with the 6-step wizard — name, persona, Slack app, credentials, tools, review.

Click New Agent in the sidebar to open the creation wizard. Six steps, ~5 minutes if your Slack credentials are already in hand.

Before you start, skim Concepts — What Goes Where. It explains the four layers of an agent’s brain (system prompt, skills, memory, wiki) so you pick the right home for each piece of content later.

Step 1: Name & Role

The agent’s identity card. Name is what teammates @mention in Slack; the model card you pick decides the cost/quality tradeoff for every reply.

Field	What it is
Agent Name	Display name shown in Slack and the dashboard (e.g. “Nelson”, “Data Analyst”)
Model	Choose between Opus 4.6 (most capable, slowest), Sonnet 4.6 (balanced — recommended default), or Haiku 4.5 (fastest, cheapest, best for high-volume narrow tasks)
This agent is a Boss	Toggle on if this agent orchestrates others. Boss agents auto-generate their CLAUDE.md from the team registry — see Boss Agents
Reports To	Only shown for specialists (Boss off). Pick one or more boss agents this specialist answers to — multi-select is supported
Tags (optional)	Free-form labels (e.g. Product, Data, Engineering) used by the dashboard’s filter chips and by Boss for routing

Start with Sonnet 4.6. Move to Opus only if the agent needs deep multi-step reasoning; move to Haiku only if you’ve measured Sonnet to be unnecessarily slow for the workload.

Step 2: Persona

Pre-built starting points across Engineering, Data, Product, Business, and Generic categories. Use the search box to find one fast, or click Blank to start from scratch.

Picking a persona seeds the agent’s:

• System prompt base (role, tone, hard rules)
• Starter skills (the per-card “5 skills” badge tells you how many)
• Recommended MCPs for that role

The selected card shows a green check; you can change your mind by clicking another card before continuing. See Personas for the full library and how seeding works.

Boss agents skip this step. Their system prompt is auto-generated from the team registry and cannot be hand-edited.

Step 3: Create a Slack App

Each agent needs its own Slack app. SlackHive generates a complete manifest (named {agent-slug}-manifest.json) so you don’t have to hand-author scopes, event subscriptions, or socket-mode config.

The on-screen flow is:

1. Open api.slack.com/apps — click Create New App → From an app manifest, select your workspace
2. Paste the manifest — switch to the JSON tab in Slack’s dialog, paste, click Next → Create
3. Install to workspace — sidebar Install App → Install to Workspace → Allow. The Bot Token only appears after this install step.

Click Copy manifest in the wizard to grab the JSON, then paste it into Slack. See Slack app setup for screenshots of the Slack admin side and common troubleshooting.

Step 4: Credentials

Three values from your fresh Slack app. Each field has a Show guide link that expands an inline screenshot of the exact Slack admin screen — no need to context-switch back and forth.

Credential	Slack location	Format
Bot Token	OAuth & Permissions → Bot User OAuth Token	xoxb-...
App-Level Token	Basic Information → App-Level Tokens → Generate Token and Scopes, add connections:write scope	xapp-...
Signing Secret	Basic Information → App Credentials → Signing Secret → Show	32-char hex

The Continue button stays disabled until all three fields are populated. Credentials are encrypted at rest in the platform integrations table — separate from the agent row itself. One agent can later hold creds for additional platforms (Discord, Telegram) as adapters land. See Platform integrations for rotating tokens after save.

Step 5: Tools

Pick the MCP servers this agent should be able to call — Notion, Jira, GitHub, Datadog, your custom internal MCPs, anything in the catalog. Each row shows a one-line description and a transport badge (stdio, sse, http).

You can:

• Skip this step entirely — assign MCPs later from the agent’s Tools tab. Nothing breaks.
• Tick what you need — assignments are saved with the agent on Create Agent.

Capability toggles like Internet Access and Shell Access live on the post-creation Tools tab, not in this wizard. You manage those alongside MCPs after the agent is up. See Tools and MCP servers.

Step 6: Review

A read-only summary of everything you entered: name, model, role, reports-to, description, masked tokens, MCPs, and the seeded skill template. The green callout sets the right expectation:

Once created, the runner picks up the agent automatically and connects to Slack. Manage skills, MCPs, and channel permissions from the agent detail page.

Click Create Agent to launch. The dashboard status flips to Running within a couple of seconds — that’s the runner spawning the Claude session and opening the Slack socket connection.

After creation

The agent is live. Invite the bot to a Slack channel:

/invite @your-agent-name

Then @mention it or DM it directly. First reply takes a couple seconds as the session initializes.

The agent detail page

Open the agent to find five tabs:

Tab	What you manage here
Overview	Name, description, model, persona, Slack credentials, channel restrictions, verbose toggle (see below)
Instructions	System Prompt, Skills, Memory (three sub-tabs)
Tools	MCP servers plus Internet Access and Shell Access toggles
Wiki	Assign shared wiki folders from the Knowledge Library
Test	Sandboxed conversation pane with real MCPs
History	Auto-snapshots with GitHub-style file diffs and restore

Changes hot-reload within seconds - no restart needed.

Instructions tab - three sub-tabs

Sub-tab	What lives here	Links
System Prompt	The agent’s identity. Always-loaded. Edit by hand or via Coach. Optimize button runs Coach to tighten it.	Concepts
Skills	Reusable slash-command procedures. Each is one Markdown file.	Coach
Memory	What the agent has learned from conversations. Inlined into the system prompt every turn.	Memory

Boss agents: the System Prompt is locked and the Optimize button is hidden. Boss identity is auto-generated from the team registry whenever any specialist is added, updated, or deleted.

Skill descriptions

Each skill carries a short one-line description that gets inlined into the agent’s compiled CLAUDE.md as a “when to use” hint — so the model can pick the right /command without loading the full body. In the Skills sub-tab, click into a skill to see a small grey DESC chip next to the filename. Click the chip to open the edit modal:

• Save — write a manual description (free-form text, ~80 char target).
• Regenerate with AI — ask the runner to summarize the skill body via Claude Sonnet 4.6 into a “Use when…” trigger phrase. The modal stays open with a spinner overlay until the new draft arrives, then auto-populates so you can review before saving.

Descriptions auto-fill on first save (no human input needed) and self-heal on every runner restart — any skill row that’s still NULL gets summarized in the background sweep. Empty descriptions render the skill as - /name in the index; populated ones render as - /name — Use when ….

Tools tab

The Tools tab has two sections:

• MCP servers - assignment grid for servers in the catalog
• Capabilities - toggles for Internet Access (web search + fetch) and Shell Access (terminal commands, dangerous commands auto-blocked)

Read and Write are always on (the memory system needs Write). See Tools.

Verbose toggle

A per-agent switch on the Overview tab controls how chatty the agent is while working.

Setting	Behavior
Verbose on (default)	Each assistant text block posts to Slack as it streams. Users see step-by-step progress - great for long tool chains.
Verbose off	Intermediate blocks are buffered silently. Only the final answer posts. Cleaner for short-answer agents (e.g. @status-bot).

Tool progress always shows as emoji reactions on the user’s message regardless of this toggle - verbose only controls intermediate text messages.

Channel restrictions

By default, an agent responds in any channel it’s invited to. To restrict:

1. Open Overview
2. Scroll to Allowed Channels
3. Enter one or more Slack channel IDs (C12345678)

If the agent is invited to a non-allowed channel, it posts a notice and leaves.

Editing is live

Every save to the system prompt, skills, tools, permissions, or wiki:

• Hot-reloads the agent within a couple seconds
• Creates a snapshot (see Version History) - skipped if nothing actually changed
• Channel restriction changes appear in the diff

Next steps

Coach

Tune the system prompt and skills via chat - no hand-editing.

Knowledge Library

Build shared wiki folders from repos, files, or URLs and assign them to this agent.

Test Mode

Try the agent in a sandbox before going live on Slack.

Tools & Permissions

Full reference for MCPs, Claude tools, allowlist/denylist, and namespace-scoped Bash.