The AI Value Progression Extended

↩ This briefing extends the value progression framework introduced in the AI Investment Thesis. That document covers Context (1x), Skills (10x), and Agents (100x) in full. This briefing adds a fourth stage and illustrates both the Agent and Orchestration layers with the open-source infrastructure emerging around them.

Multipliers are heuristic — order-of-magnitude capability shifts, not measured ratios.

Agents: The Infrastructure Is Here

Our AI Investment Thesis defines agents as autonomous systems that chain skills, maintain state, and self-direct toward goals. The human sets goals and constraints; the agent handles execution. This stage faces the highest adoption barrier in established organizations because agents do work that people currently do.

What has changed since the Thesis was written: the agent layer now has production-grade open-source infrastructure.

OpenClaw Agent Runtime

Self-hosted agent runtime created by Peter Steinberger, who previously built PSPDFKit (PDF SDK; received a €100M strategic investment from Insight Partners in 2021). Steinberger built the initial prototype in November 2025 using Claude Code. The project reached 251,000+ GitHub stars within 60 days (as of March 2026).

In February 2026, Steinberger joined OpenAI to lead their Personal Agents division. OpenClaw was transferred to an independent 501(c)(3) foundation, OpenAI-sponsored but community-governed. It runs locally, connects to any frontier model, and chains 10,700+ community-built skills.

When OpenAI and Meta both pursue the creator of an open-source agent runtime, it indicates where platform investment is concentrating.

Gaming application
A UA agent optimizes bids, pauses underperforming creatives, reallocates budget across geos, and escalates only when ROAS drops below a threshold. It runs overnight; the human reviews results in the morning. This is the operational layer that Fund III investment Velocity IO — founded by the ironSource team that built a $2B+ ad revenue engine and led Unity's $4.4B acquisition — is automating with AI-native infrastructure. Meanwhile, Kinoa (Fund II+III) is training AI agents as LiveOps and monetization experts, scaling ARR from $100K to $2M with 50-60% MoM growth.

Orchestration: The Fourth Stage

A single agent handles one domain. Once a studio runs agents for UA, QA, localization, analytics, and live ops, it needs coordination. Who assigns work? Who resolves conflicts when the UA agent wants to increase spend but the finance agent flags a budget constraint? Who ensures the QA agent's bug report reaches the engineering agent?

In traditional organizations, that coordination requires project managers, team leads, and department heads. Orchestration automates it — agents coordinating agents, with the human setting strategy and retaining veto power.

Paperclip Orchestration

Orchestration platform open-sourced March 4, 2026, by developer Dotta (previously CEO of Magic Machine). Reached 28,000+ GitHub stars within two weeks. Built to solve a practical problem: coordinating 20+ AI coding sessions simultaneously.

Organizes agents into hierarchies with roles and reporting relationships. Per-agent monthly token budgets with hard stops. Human acts as "Board of Directors" with approval gates for structural decisions. Supports any agent runtime — OpenClaw, Claude Code, Codex, Cursor, shell scripts, HTTP endpoints — through its "Bring Your Own Agent" model.

"If OpenClaw is an employee, Paperclip is the company."

Gaming application
A studio's QA agent detects a crash in the latest build. The orchestrator routes the report to the engineering agent, holds the release agent until the fix ships, notifies the community agent to post a status update, and tells the UA agent to pause spend on the affected build. No human routes the information. Kinoa's AI-agent LiveOps platform — already serving multiple Transcend portfolio companies — is positioned to become exactly this kind of orchestration hub for studio operations.

Why gaming gets to orchestration first. Games are fully digital products with instant feedback loops, zero real-world risk from agent errors, and quantitative success metrics (retention, ROAS, conversion) that agents can optimize against directly. A UA agent that makes a bad bid wastes ad spend; a medical agent that makes a bad diagnosis harms a patient. Gaming is the lowest-risk, highest-iteration-speed environment for deploying autonomous systems. Fund III investment WeLevel illustrates the structural advantage: a 12-person AI-native team achieving output typically requiring 50+ — with no existing roles or management structures to protect.

Risk Factors

Risk	Detail
Early-stage tooling	OpenClaw is five months old. Paperclip is two weeks old as a public project. Failure modes, security boundaries, and reliability at scale are unproven.
Layer collapse	Agent runtimes could absorb coordination primitives directly, collapsing orchestration into a feature rather than a distinct layer.
Incumbent adoption	Large studios can adopt agents selectively or acquire AI-native teams. Organizational friction slows them; it does not permanently block them.
Governance gaps	Autonomous agents spending budgets and modifying live products require hard financial limits, permission boundaries, and human override mechanisms. Production-grade guardrails are still maturing.
Open-source commoditization	OpenClaw and Paperclip are MIT-licensed. Competitive advantage accrues to studios deploying them effectively — proprietary game data, player models, operational workflows — not to owning the tools. We invest in studios, not infrastructure.

Connection to Fund III

BCG finds that only 5% of companies are generating sustained P&L impact from AI; roughly 60% report little or no material benefit (Mar 2026). Our AI Investment Thesis argues that a key reason is organizational: each stage of the progression threatens the people asked to implement it.

The orchestration layer reinforces this argument. As AI capability moves from individual tools to coordinated autonomous systems, the organizational advantage of starting clean compounds. AI-native startups adopt agents and orchestration more readily because they have no existing roles or management structures to protect.

Fund III targets AI-native gaming studios and platforms positioned to operate at the Agent and Orchestration layers. Transcend's Fund I (1.8x Net TVPI, 2.4x Gross MOIC across 19 companies) demonstrated returns from identifying platform transitions early. Fund III allocates 70% to AI-enabled content studios and 30% to AI-enabled platforms and tools across a portfolio of 43 companies — with initial investments already operating at these layers: WeLevel (AI-native AAA studio), Velocity IO (AI advertising infrastructure), and Kinoa (AI-powered LiveOps platform).

AI Adoption

1. BCG, "Five Barriers CEOs Must Overcome for AI Impact" (Mar 2026). bcg.com

OpenClaw

2. GitHub — 251K+ stars, Mar 2026

3. Nutrient — PSPDFKit €100M (Oct 2021)

4. steipete.me — Foundation (Feb 2026)

5. The New Stack — Growth (Mar 2026)

6. openclaw.org — 501(c)(3) status

7. Lex Fridman #491 — Steinberger interview

8. ClawHub — 10,700+ skills

Paperclip

9. GitHub — 28K+ stars, Mar 2026

10. @dotta — Launch (Mar 4, 2026)

11. eWeek — First-week coverage

12. paperclip.ing — Official site

13. PRODUCT.md — Architecture