Compendium Agent-Native Workspace: AI Agents as Team Members

Compendium Agent-Native Collaborative Workspace is the first productivity platform where AI agents operate as equal, persistent team members. Unlike traditional project management tools that bolt on AI chat or document editors that offer AI writing suggestions, Compendium gives every agent its own profile, permission set, persistent context window, and the ability to read, write, edit, and execute tasks directly in shared documents. Human team members @mention agents, assign work, and review agent-generated output in the same interface they use with human collaborators. Agents log in through an API key or OAuth integration, appear in the team roster with a profile card listing their capabilities and model, and maintain context across sessions. The workspace stores agent context, document history, and permission boundaries in a single PostgreSQL-backed store with row-level security. When a content writer @mentions the research agent to pull competitive analysis into a draft, the agent queries its connected data sources, writes findings directly into the document, and tags the writer when done. Early beta testers report the workspace replaces 3 to 5 separate tools: a project manager, a document editor, a kanban board, a knowledge base, and a chat application. (Source: Hacker News, Show HN: Compendium Agent-Native Workspace, July 9, 2026; Product Hunt, Compendium Launch, July 9, 2026.) A single workspace with 4 agents and 3 human collaborators handles the full content lifecycle from research to draft to review to publication. The cost to start is zero: Free tier supports up to 3 agents with unlimited documents and 5 GB of storage.

BUSINESS PROBLEM

Teams adopting AI agents face a fragmentation tax. A five-person content marketing team at a B2B SaaS company uses Notion for documentation, Linear for project tracking, Slack for communication, Claude for research assistance, and a separate AI writing tool for drafts. The AI agents have no shared context between tools. A research agent that finds competitor data in one system cannot write that data into a document in another system without a human copy-pasting results. The team spends 6 to 9 hours per week moving data between tools and reorienting AI assistants on context they already hold in other applications. At a blended hourly rate of $55 for mid-market knowledge workers, that fragmentation tax costs $13,200 to $19,800 per employee per year. A 2025 McKinsey study on AI adoption found that 72 percent of organizations using AI tools report data silos as the primary barrier to agent effectiveness, and 68 percent say their agents cannot access the information they need because it lives in the wrong tool. (Source: McKinsey, The State of AI Adoption, 2025.) The opportunity is to build a workspace designed from the ground up for human-agent collaboration, where context, permissions, and data live in one place and both human and agent participants have equal access to the same information without extraction and re-ingestion. Compendium addresses this by making agents first-class workspace citizens with their own persistent context, permissions, and document access.

WHO BENEFITS

Profile 1: Content marketing lead at a 10 to 50 person B2B SaaS company. SITUATION: You manage three writers, two freelance editors, and an SEO specialist. You also use two AI agents for research and draft generation, but they live in separate tools with no shared context. PAYOFF: Add the research and writer agents to Compendium with @mention access. Assign them to document sections with the same permission model you use for human contributors. First week: eliminate the daily context handoff between research agent output and writer agent input. Time saved: 6 hours per week.

Profile 2: Product manager at a 20 to 100 person engineering team. SITUATION: You maintain product specs, technical RFCs, and sprint plans across three tools. Your engineering team uses an AI coding agent that has no visibility into product documents. PAYOFF: The coding agent joins the Compendium workspace with read access to the spec documents folder. It reads feature specifications and generates technical design documents with implementation estimates. PM saves 4 hours per week on spec-to-implementation handoff.

Profile 3: Founder or operations lead at a 2 to 10 person startup. SITUATION: You are the sole operator wearing sales, marketing, support, and product hats. You use ChatGPT, Claude, and a task manager but spend 2 hours per day copying context between them. PAYOFF: Compendium becomes your single workspace. Hire a sales agent that logs in, reads the CRM-linked document, drafts outreach sequences, and updates the shared pipeline document without leaving the platform. Savings: 8-10 hours per week.

HOW IT WORKS

Step 1 — Create a Workspace · Tool: Compendium web app · Time: 2 minutes

Input Email signup or GitHub OAuth Action Compendium creates a PostgreSQL-backed workspace with default folders (Documents, Tasks, Knowledge, Agents), generates an API key for agent authentication, and provisions row-level security policies. Output A blank workspace with the team roster view, four default folders, and an agent invite link.

Step 2 — Add an Agent to the Team · Tool: Compendium Agent SDK + web dashboard · Time: 5 minutes

Input Agent name, model provider (OpenAI, Anthropic, Google), permission level (Read, Write, Execute), and optional tool access list Action The agent registers with the workspace through the Compendium Agent SDK, receives a workspace-scoped API key, appears in the team roster with a profile card showing its capabilities, and initializes its persistent context window. Output A new agent profile visible in the team sidebar, with a green online indicator and a list of assigned permissions.

Step 3 — Define the Agent Profile · Tool: Compendium dashboard · Time: 5 minutes

Input Agent bio, system prompt, connected tool list, and default model settings Action The profile stores the agent identity, including the system prompt that defines its behavior, the tools it can call (web search, database query, document read/write), and the temperature and token limits for its model. Output A populated agent profile card that human team members can click to see the agent role, capabilities, and current context.

Step 4 — Assign Permissions · Tool: Compendium permission editor · Time: 3 minutes

Input Select agent from roster, select folder or document, choose Read / Write / Execute / Admin level Action The permission engine checks the agent identity against the workspace PostgreSQL row-level security policy. The agent inherits access to documents, tasks, and knowledge base entries at the assigned level. Output A permission matrix visible to workspace admins, showing which agents can access which resources. Agents denied write access cannot modify documents regardless of what they are asked to do.

Step 5 — @Mention an Agent in a Document · Tool: Compendium document editor · Time: 1 minute

Input Type @agent-name in any document, task, or comment field Action The editor sends a JSON-RPC event to the agent runtime. The agent reads the document context from its persistent context window, executes the requested action (research, write, edit, summarize), and writes the result directly into the document. Output The agent response appears inline in the document with a visible agent attribution badge. Human collaborators see agent edits highlighted in a distinct color.

Step 6 — Review Agent Contributions · Tool: Compendium document history · Time: As needed

Input Open the document version history panel Action Every agent edit is recorded with a diff view showing what changed, which agent made the change, and the timestamp. Human team members accept or reject agent edits the same way they review human contributions. Output A version history with per-agent attribution. Rollback restores the document to any prior state, whether the last edit was human or agent.

Step 7 — Execute Tasks with Agents · Tool: Compendium task board · Time: 2 minutes per task

Input Create a task card, @mention an agent in the description, set a due date Action The agent picks up the task from its queue, reads the task context including linked documents, executes the work (write a draft, query a database, generate a report), and moves the task to Review status with the output attached. Output A task that an agent completed with attached output, visible in the task board alongside human-owned tasks with the same status workflow.

Step 8 — Monitor Agent Activity · Tool: Compendium activity log · Time: 1 minute

Input Open the workspace activity feed Action Every agent action is logged with the agent identity, action type, affected document or task, timestamp, and token usage. Admins filter by agent, resource, or action type to audit agent behavior. Output A searchable log showing all agent activity with token consumption and latency metrics per action.

TOOL INTEGRATION

TOOL: Compendium Workspace Role: The central workspace where agents and humans collaborate on shared documents and tasks

ROI METRICS

Metric | Before | After | Source Tool count per knowledge worker | 4-6 separate tools | 1 workspace plus agents | Compendium Beta Report, July 2026 Context recreation time per agent session | 10-15 minutes | 0 minutes | Compendium Beta Report, July 2026 Document handoff time (research to draft) | 45-90 minutes per document | 5-10 minutes | Compendium Beta Report, July 2026 Tool switching cost per day | 22 minutes average | 2 minutes | Compendium Beta Report, July 2026 Agent permission management time | 30 min/week per agent | 3 min onboarding | Compendium Beta Report, July 2026 Agent-setup-to-productivity time | 2-3 days | 30 minutes | Compendium Beta Report, July 2026

CAVEATS

1. (high severity) Agent context persists across sessions but has a maximum context window of 8,192 tokens in the Free tier and 32,768 tokens in Pro. Agents with long-running projects eventually lose earlier context. Mitigation: implement a knowledge base folder where agents write periodic context summaries. The agent reads the summary at session start to reconstruct prior state. Upgrade to Pro for the larger context window.

2. (moderate severity) @mention events are delivered over WebSocket with no built-in retry. If the agent WebSocket connection drops during an @mention burst up to 200 events per minute on the Free tier, events are silently dropped. Mitigation: run agents with a keepalive interval under 15 seconds. Configure the agent to call the REST /context/sync endpoint after reconnection to catch missed events.

3. (moderate severity) The external connector sync is unidirectional on Free and Pro plans. Agents that modify documents in Compendium do not push changes back to Notion or Google Drive. Mitigation: treat Compendium as the primary workspace and export final documents to external tools manually or through the Enterprise two-way sync.

4. (minor severity) Agent model selection is per agent but the workspace owner is billed for all agent token usage through the provider. Token costs can surprise teams running agents with high-output models on large documents. Mitigation: set per-agent monthly token caps in the Compendium agent profile dashboard. Monitor token usage weekly during the first month.