Agents

Agents are pre-built workflows your team can call from any chat. Type @meeting-recap in LibreChat and an agent searches the last 7 days of meetings, drafts a recap, asks an admin to approve key facts, and saves the result to your team's memory — all without you leaving the chat.

GrooveOS ships a starter agent library curated by the GrooveOS team. Your team admins can edit, disable, or add new agents from the admin panel. Every agent runs inside your team's scope — it reads only your team's memory, writes only to your team's memory, and shows up in your team's audit log.

What agents do

Each agent is a small workflow built on top of LangGraph (the same open-source agent framework used at OpenAI, LangChain, and others). The workflow can:

• Search your team's memory, CRM, tasks, meeting notes, or Drive-synced docs.
• Call any MCP tool your team has enabled (scrape a URL, read calendar, etc.).
• Use Claude, GPT-4, or Grok — the agent picks the right model for each step.
• Pause and ask a team admin to approve a step before continuing (HITL).
• Write the result back into memory at the truth level chosen by the admin.

Because the agent runs inside your team's scope, the same agent gives Team A and Team B different results: the inputs (your memory) and outputs (your audit log) are isolated end to end.

Invoking an agent from chat

The simplest way to run an agent is from LibreChat:

1. Open chat.grooveos.app and pick any conversation (or start a new one).
2. Type @ — an autocomplete menu lists the agents your team has access to.
3. Pick one — for example @meeting-recap.
4. Add any free-form prompt the agent needs (e.g. "for the last week").
5. Press Enter. The agent runs in the background; progress streams into the conversation.

You: @meeting-recap for the last 7 days
@meeting-recap: Reading 4 Granola meetings + 12 memory items...
@meeting-recap: Found 3 candidate facts. Sending to admin for approval.
@meeting-recap: Approved by Alice. Recap saved as memory_item mem_xyz...

The starter agent library

GrooveOS ships with these agents pre-seeded on every new team:

Your team's admins can add their own agents in the admin panel. See Admins — curating the agent library below.

Approvals — human in the loop

Agents are not autonomous black boxes. When an agent reaches a low-confidence decision — for example, it isn't sure whether a fact deserves VALIDATED or CANONICAL — it pauses and asks a team admin for approval.

The admin sees the question in their notification feed:

@meeting-recap is paused — needs your decision

Fact: "Q2 fundraising target is 2M€"
Confidence: 0.62
Suggested truth level: WORKING
Source: Granola note "Investor call — May 6"

[ approve as WORKING ]  [ promote to VALIDATED ]  [ discard ]

The admin clicks one button. The agent resumes from exactly where it paused and keeps going — even if the approval comes hours or days later. Nothing is lost between pause and resume.

Admins — curating the agent library

Team admins manage the agent library from grooveos.app/account/admin/agents/. From there you can:

• Enable or disable any starter agent for your team.
• Edit the prompt or model picked by a starter agent.
• Add a new agent definition — give it a name, a prompt, and a list of MCP tools it may call.
• Inspect every past run (inputs, intermediate steps, final output, admin approvals).

Agent definitions live in the agent_definitions table inside your team's scope. Edits propagate immediately — the next time a member types @your-agent they get the new prompt and tool list.

# Example agent_definition row (read-only view)
{
  "name": "weekly-digest",
  "description": "Compiles last week's decisions for stand-up.",
  "model": "claude-sonnet-4-7",
  "tools": ["memory_search", "task_list", "contact_lookup"],
  "system_prompt": "You are an internal scribe. Summarize last week...",
  "requires_admin_approval": true,
  "truth_level_on_success": "WORKING"
}

Where agent output lands

Every agent run produces three artefacts you can browse later:

• A conversation in the chat where you invoked it — readable from chat.grooveos.app.
• Memory items the agent wrote — readable from the Brain Monitor and searchable via the API.
• An audit-log entry per write — admins can see who ran the agent, what it wrote, and which approvals it asked for.

The same audit feed shows up in the superadmin dashboard at grooveos.app/account/admin/ so the team owner can confirm agents are behaving as intended.

FAQ

Are my prompts to an agent shared with other teams?

No. Agents run inside your team's scope. Your prompts, your team's memory, and the agent's writes are isolated end to end — Team A's @meeting-recap cannot see Team B's meetings.

Can I cancel a running agent?

Yes. The chat conversation shows a Stop action while the agent is mid-run. Stopping mid-run leaves no partial writes — any writes that were not already approved are rolled back.

How much does an agent run cost?

Agent runs draw from your team's plan-included LLM credits (Team plan: 5M tokens / month, Enterprise: custom). Heavy users on the Team plan can buy add-on token packs.

What's next

• Memory & Knowledge — How agents write into the truth-level pipeline.
• Brain Monitor — Where to see what your agents have written.
• MCP Tools — The tools your agents can call.
• API Reference — Invoke agents programmatically.