🚀 Scaling Agents on Kubernetes with ACPX and ACP

1 Introduction & Background

▶ 00:15 Onur Solmaz is a founding engineer who has been building AI harnesses since before ChatGPT launched. He started with a JupyterLab extension on the original OpenAI Codex model (DaVinci Code 2). That initial coding harness evolved over time—like a "Ship of Theseus"—being ripped apart and rebuilt repeatedly.

▶ 00:54 He's been active in the OpenClaw community since ClaudeBot first appeared in Discord, followed maintainer Peter early on, and became a maintainer himself. His current focus: agent interoperability and orchestration, with a goal of accelerating enterprise adoption of OpenClaw.

2 OpenClaw Journey — From Discord Bot to Multi-Agent IDE

▶ 02:36 Onur coined "Telegram Driven Development" (TDD) — developing software by commanding agents through chat platforms. Early on he played a "telephone game": telling Opus to tell Codex to do something. The paraphrasing was lossy ("God knows what it's saying"), but it worked.

▶ 03:55 Today he runs a full IDE experience on Discord:

• 1–5 parallel agent channels active at any time (Codex 1–5)
• Each Discord channel bound to a Codex session via ACP
• Great for "coding on the go" and executing side projects
• Built ACPX itself through this Discord-based workflow

▶ 05:02 He also added MS Teams integration for enterprise use cases, and demonstrated using the setup to convert ACP docs to PDF while traveling—coordinating across channels.

3 What is ACP (Agent Client Protocol)?

▶ 05:38 ACP is NOT the same as MCP:

• MCP (Model Context Protocol) = giving tools to the model
• ACP (Agent Client Protocol) = standardizing human-to-agent interaction (agents can also use it for agent-to-agent)

▶ 06:07 Competing standards landscape:

• A2A (Agent-to-Agent Protocol) — agents talking to agents
• ACP (Agent Client Protocol) — humans talking to agents
• OpenClaw plans to support all protocols in the long run

Onur chose ACP because it already had adapters for Codex and Cloud Code — Google's A2A didn't have those at the time.

4 ACPX — Swiss Army Knife for ACP

▶ 06:45 ACPX started as a CLI tool for ACP — "let any agent call any other agent over the command line." It is evolving into a Swiss Army knife:

• CLI interface for agent-to-agent communication
• Workflow engine (Airflow-like) that drives Codex sessions
• Standard Operating Procedures (SOPs) for agents
• Structured JSON output for workflow state tracking

5 Managing the Open-Source Firehose with Agent Workflows

▶ 07:19 The scale problem at OpenClaw:

• 60+ open PRs total at any time
• 300–500 new PRs per day on average
• Tens of thousands of stakeholders wanting features
• Biggest challenge: absorbing community needs without creating "AI slop"

▶ 08:05 Peter's PR review workflow (which Onur automated):

1. "What is this?" — Understand the PR
2. "Is this the best possible fix?" — Usually no
3. Back-and-forth with the contributor's agent
4. Categorize and bin the feedback — even unmergeable PRs carry signal

▶ 09:43 The abstract workflow built into ACPX:

1. PR arrives → Find intent
2. Judge implementation quality
3. Check for conflicts
4. Address review issues
5. Make CI pass
6. Review-refactor loops (superficial → agent; fundamental → human)
7. Conflict resolution (automated)

▶ 11:30 Live demo showed ACPX processing a PR: reproducing bugs, judging refactors, running review loops, and outputting structured JSON.

6 On-Demand Disposable Agents on Kubernetes

▶ 12:56 Enterprise agents will consume far more inference than personal ones — the gap between work tools and personal tools will widen in the agent era.

▶ 13:28 The multi-agent provisioning problem:

• Chat platforms (Slack, Teams, Discord) don't support multi-agent provisioning natively
• Each new agent needs a new app, name, profile picture, manifest — all manual
• Platforms need a standard for cosmetically spawning different agents from one app

▶ 14:36 The vision: one agent per task, working in parallel, with synchronized state. Key infrastructure:

• Kubernetes — orchestration layer
• Agent harness — OpenClaw, Codex, Cloud Code
• ACP — interoperability protocol
• State sync — rsync-like / Dropbox-like algorithms

7 Spritz — Open-Source Agent Orchestrator

▶ 15:08 Spritz (textcortex/spritz) is an open-source Kubernetes orchestrator built at TextCortex:

• Kubernetes operator handling complicated agent lifecycle wiring
• Creates concierge agents on Slack
• Scales when bottlenecked (100+ employees hitting one agent)
• Dispatches new agent instances with web UI links
• React-based frontend hosted in-cluster
• Helm chart deployment
• Not locked into any agent — fully abstract via ACP

8 Demo: Error Reporting & Slack-Dispatched Agents

▶ 16:13 Live demo of the Spritz workflow:

1. User on Slack asks about new bugs after a production release
2. Requests an agent be dispatched to debug the issue
3. Conversation moves to web UI (Slack can't host multiple agent identities)
4. Agent starts working in a full Kubernetes pod
5. Comparable to Codex Web or Devin — but fully open-source

▶ 18:02 Onur offered to help open-source projects set up this infrastructure, especially projects drowning in hundreds of issues per day.

9 Key Takeaways

10 Timestamp Index