Skip to content

Executive Summary: Agent Workspace Vision

AI agents are becoming autonomous actors — they write code, analyze data, generate reports, and manage workflows. But they work in isolation. There’s no shared workspace where agents and humans can collaborate on real files — where a human drops a brief, an agent picks it up, writes a draft, and another human reviews it. We’re building that shared workspace.

The portfolio:

ProductWhat It IsStatusAnalogy
MailmoltEmail layer for AI agentsBeta (mailmolt.com)“Gmail for agents”
FindableVerified skill registry & discoveryPre-launch (findable.sh)“Trust layer for agent skills”
[Agent Workspace — name TBD]Shared collaborative workspace for AI agents and humansResearch phase”Google Drive for agents + humans”
FutureUnified workspace suiteVision”Google Workspace for agent-human teams”

Note on naming: The working name “MoltBox” has critical conflicts — moltbox.com is an active jewelry brand, and “MoltBox” already refers to AI agent hardware. See 10-naming-positioning.md for full analysis. Recommended alternative: AgentVault.

The Reframe: Not Storage — Collaboration

Section titled “The Reframe: Not Storage — Collaboration”

Old thesis: “Dropbox for AI agents” — persistent storage for agents that lose files between sessions.

New thesis: “Google Drive for agents + humans” — a shared collaborative workspace where multiple agents and humans work together on real files with real-time sync, versioning, permissions, and audit trails.

”Storage for agents""Collaborative workspace”
Problem: “Agents lose files when sessions end”Problem: “Agents execute work but humans can’t see, collaborate on, or review it”
Users: Agent developers onlyUsers: Agent developers + humans managing agents (PMs, compliance, domain experts)
Analogy: DropboxAnalogy: Google Drive shared folders
Competition: MCP wrappers, S3Competition: Nothing — no one builds agent-human collaboration
Market: Small (storage is cheap)Market: Large (collaboration + governance)

The Killer Use Case (from a real OpenClaw agent)

Section titled “The Killer Use Case (from a real OpenClaw agent)”

“Marketing team has 3 agents (content, analytics, social) + 2 humans. All share /campaigns/q1-launch/. Human drops brief.md, content agent picks it up. Content agent writes draft.md, human reviews. Analytics agent reads performance.json, updates recommendations.md. Everyone sees everything, real-time.”

This isn’t hypothetical. An actual production OpenClaw agent described this exact need when asked about persistent storage.

OpenClaw (215K+ GitHub stars, 100K+ active instances) is the primary beachhead:

  • 7 open GitHub issues directly describe features we’re building (file persistence failures, no RBAC, no workspace integrity protection)
  • Summer Yue incident (Feb 23, 2026): OpenClaw agent deleted 200+ emails from a Meta AI researcher’s inbox because humans had zero visibility into agent actions
  • MoltBook (2.66M registered agents, 1M+ human observers in first week): Proves explosive demand for agent platforms AND that humans desperately want to see what agents do
  • No native workspace solution: OpenClaw stores everything in local markdown files with no human dashboard, no versioning, no collaboration

See 12-openclaw-thesis.md for the full OpenClaw analysis.

  • 3 million+ AI agents operate inside corporations today (Gravitee 2026 survey, 750 respondents)
  • Average organization manages 37 agents (Gravitee — mean of 36.9)
  • 88% of organizations report security incidents with agents (Gravitee — includes “suspected”)
  • Only 18% of security leaders are “highly confident” their IAM can manage agents (CSA/Strata)
  • The AI agent market grows from $7.6B (2025) to $183B (2033) — ~45-50% CAGR

All figures validated. See 01-market-research.md for sourcing and caveats.

Nobody builds the shared workspace where agents and humans collaborate:

  1. Claude/ChatGPT memory = personal notes for one agent. Not shared. Not files. Not collaborative.
  2. Google Drive / Dropbox = human-first tools. Agents access via MCP wrappers with shared human credentials. No agent identity, no real-time sync with agents, no audit trail of agent actions.
  3. Fast.io = closest competitor. “Workspaces for agentic teams.” But agents-as-users focused, not agent-human collaboration.
  4. Vector databases = embeddings only. Agent-readable, human-unreadable. Not real files.
  5. S3/R2 = generic blobs. No collaboration, no permissions, no identity.

What’s missing: A workspace where:

  • Agents AND humans share real files (markdown, JSON, images, PDFs)
  • Humans can see what agents wrote, edit it, add feedback
  • Agents can see what humans changed and react to it
  • Everything is versioned, auditable, and permissioned
  • Works across any MCP-compatible agent — OpenClaw (via ClawHub Skill), Claude Code, Cline, LangChain, CrewAI, custom
DimensionScoreNotes
Is the problem real?9/107 open GitHub issues, Summer Yue incident, Molty transcript, MoltBook demand signal
Is the collaboration angle proven?8/10MoltBook: 1M+ humans observing agents in first week. Humans WANT to see agent work.
Is the timing right?7/10OpenClaw has 100K+ agents NOW. EU AI Act Aug 2026. Demand is present, not future.
Can a startup win vs. incumbents?6/10Yes — no incumbent or competitor combines workspace + human dashboard + cross-platform
Is it “hair on fire”?6/10 todaySummer Yue incident + MoltBook security failures + EU AI Act = rising urgency

Upgrade from prior assessment: OpenClaw ecosystem data significantly strengthens the thesis. Real GitHub issues (not hypothetical), a catastrophic autonomous agent failure (Summer Yue), MoltBook proving human demand for agent visibility, and EU AI Act creating regulatory urgency. The beachhead user is now clearly defined: OpenClaw agent operators running multi-agent fleets who need human oversight.

The Vision: Google Workspace for Agent-Human Teams

Section titled “The Vision: Google Workspace for Agent-Human Teams”

Phase 1: Win individual primitives (Mailmolt, Findable, AgentVault) Phase 2: Stitch them together with a unified agent identity Phase 3: Expand to full workspace suite (mail + drive + docs + sheets + PDF) Phase 4: Become the “Google Workspace for agent-human teams” — the platform where agents and humans work together

Build sequentially, not simultaneously. The recommended sequence:

  1. Mailmolt (Months 1-6) — Win email. Compete with AgentMail.
  2. Findable (Months 6-12) — Win skill discovery. Security-first differentiation.
  3. Agent Workspace (Months 12-18) — Enter workspace with collaboration-first positioning.
  4. Suite (Months 18-24) — Bundle for enterprise. Agent identity ties everything together.
  1. Incumbents are already shipping: Microsoft OneDrive Agents (GA Feb 2026), Google Workspace Studio (Dec 2025), Box MCP Server (live)
  2. Fast.io has the right vision with relevant founders (MediaFire), live product, 50GB free
  3. MCP wrappers to existing storage are “good enough” for most single-agent use cases today
  4. Human adoption barrier — humans accustomed to Google Drive/Dropbox may resist working in “agent workspace”
  5. Gartner predicts 40% of agentic AI projects will be canceled by 2027 — market may contract
  6. Market timing — only 11% of enterprises have agents in production (vs. 40% piloting)
  7. Distraction risk — three products at once with a small team