Skip to content
Aakash GuptaAakash Gupta

How to build a Company Operating System with Hermes and OpenClaw

A CPO spent 5 months building an entire company operating system on top of OpenClaw and Hermes. It writes status reports, gatekeeps stakeholder feature requests, runs his calendar and inbox, maintains the design system in Figma, and handles 70 to 75% of the recruiting workflow. In this episode, he opens the covers and walks through the complete architecture. Full Writeup: Transcript: https://www.aakashg.com/how-a-cpo-built-a-company-operating-system-with-openclaw-and-hermes/ GitHub repo: https://github.com/mshcheglov1-ux/corporate-waters-ai-agent Timestamps 0:00 - Intro 3:55 - Inside the company knowledge graph 8:00 - How to measure product context coverage 12:08 - What the agentic CPO owns 16:52 - The Slack agent for status reports and feature requests 22:22 - The architecture using OpenClaw & Hermes 31:37 - Writing imperatives that stop fake helpful answers 37:43 - Auto-generated skills and the 31% recall jump 41:35 - The board skill, the design system, and backlog management 54:05 - Automating 70% of the recruiting workflow 57:13 - The future of PM jobs and how to hire AI native PMs 🏆 Thanks to our sponsors: 1. Bolt.new - Ship AI-powered products 10x faster - https://tinyurl.com/y8srfztt 2. Product Faculty - Get $150 off their #1 AI PM Certification: code AAKASH150 - https://www.productfaculty.com/?code=AAKASH150 3. Customer.io - Send smarter messages using your product data - http://customer.io/productgrowth 4. Land PM Job - 12-week live course to master the PM job search - https://www.landpmjob.com/ 5. Viktor - Use $100 in starting credits to get 5x more done with this AI employee - http://ref.viktor.com/aakashgupta3 Key Takeaways: 1. The better your AI knows your context, the higher-level work you can delegate to it. Usually, one person builds deep context in a specific domain and then leaves, taking it all with them. A single store of business, customer, product, and technical knowledge removes that bottleneck. 2. Product context coverage is a KPI you can track. The agent is asked what percentage of knowledge it holds across industry, business model, and customers, with each term defined precisely. 3. Never summarize what you feed the agent. Summarizing loses granular detail and forces every conversation into a fixed template. Testing showed roughly 20 to 25% worse recall on summarized inputs. Store raw transcripts instead. 4. Build memory in 3 layers - Layer 1 is the knowledge graph of interconnected entities. Layer 2 is a vector database, because most requests are fuzzy and keyword matching fails around 75% of the time. Layer 3 is every conversation, meeting, and agent reflection written to raw MD files daily. 5. Blend OpenClaw with Hermes. OpenClaw gives you scaffolding that works out of the box with strong engineering support. Hermes adds automatic skill generation based on the tasks you request most often. Tested across 5 core topics, the auto-generated skills improved recall by 31%. 6. Make the agent the stakeholder gatekeeper. Stakeholders are trained to bring feature requests to the agent before they reach a PM. The agent asks clarifying questions, checks the request against existing priorities, and declines it politely if it does not hold up. If it does hold up, the agent adds it to the backlog and routes it to the domain owner it already has mapped. 7. Turn your board into a skill. Board meeting transcripts let the agent abstract each director's mental model into a set of principles. Before defending a strategy, run the deck through that skill and ask it to poke holes. You get the brutal version of the feedback before the room does. 8. Automate the recruiting funnel end to end. LinkedIn Recruiter handles sourcing and templated outreach, the ATS is driven entirely through the agent, and transcribed interviews produce a third-person read on every candidate. That covers 70-75% of the workflow. The rejection emails come back with specific improvement areas pulled from the transcript, which most recruiters never do. 9. Ask candidates what they have already automated. The AI qualifier is one question about which parts of their daily work they have automated. Answers range from chatting with a web interface to running a full agent stack, and that spread tells you everything. 👨‍💻 Where to find Mikhail Shcheglov LinkedIn: https://www.linkedin.com/in/scheglovm1/ X: https://x.com/mshcheglov1 Substack (Corporate Waters): https://corpwaters.substack.com/ 👨‍💻 Where to find Aakash: X: https://x.com/aakashgupta LinkedIn: https://www.linkedin.com/in/aagupta/ Newsletter: https://www.news.aakashg.com #ProductManagement #AIProductManagement 🧠 About Product Growth: The world's largest podcast focused solely on product + growth, with over 200K+ listeners. 🔔 Subscribe and turn on notifications.

Aakash GuptahostMikhail Shcheglovguest
Jul 31, 20261h 6mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Building an AI chief of staff through knowledge graphs and agents

  1. OLX built a company knowledge graph over five months and tracks “product context coverage” (e.g., 54%) as a KPI for how well the AI understands the business, customers, and market to safely increase autonomy.
  2. The “agentic CPO” shifts from process/scaffolding ownership to owning an AI operating system—architecture, rituals, rules, access control, and continuous iteration—so AI can handle repetitive work and elevate PM leverage.
  3. A Slack-based agent handles status reporting, stakeholder feature-intake triage, and personal exec assistance (email/calendar digests and scheduling) using a large library of imperatives to reduce hallucinations and “fake helpful” answers.
  4. The underlying stack blends OpenClaw (scaffolding/tools) with Hermes (auto-generated skills), uses three persistent memory layers (knowledge graph, vector DB, raw transcripts), and applies hybrid retrieval to avoid context overload.
  5. Agents extend into high-value workflows like board-deck critique (“board skill”), design system generation/maintenance, backlog abstraction across spreadsheets, and recruiting automation covering ~70–75% of the funnel.

IDEAS WORTH REMEMBERING

5 ideas

Treat knowledge retention as a core AI-native goal, not a byproduct.

Shcheglov frames AI-native teams as a way to prevent knowledge leakage when experts leave by centralizing business, product, and customer context into a persistent, queryable system.

Measure AI usefulness with a context coverage KPI tied to delegation potential.

OLX prompts the system to estimate its knowledge across industry, business model/PNL drivers, and customer segments, using the resulting percentage as a directional indicator for how autonomously the agent can operate.

The CPO’s new leverage is owning the agentic scaffolding and rituals.

Instead of primarily enforcing planning cadences and review ceremonies, the agentic CPO builds the environment where AI collaborates with teams, makes bounded decisions, and continuously improves through fast feedback-to-deploy cycles.

Automate PM “ritual work” first to double down on discovery.

He estimates ~50% of PM time historically goes to reporting and repetitive coordination; delegating that to agents can let one PM cover the output of two while focusing human time on customer discovery and judgment.

Raw transcripts can outperform summaries for retrieval quality.

They found summarization degraded recall (20–25% worse) by removing nuance and forcing templates, so they store full transcripts and rely on retrieval to pull only relevant slices.

WORDS WORTH SAVING

5 quotes

the core goal of having AI native teams is to be able to automate as much knowledge as we can.

— Mikhail Shcheglov

it understands our industry, uh, like classifieds, our market, our business, our business model specifically, the top line, the bottom line, the drivers behind it, our customers and the value drivers for them at the level of 54%, which already means that it can operate as a capable, I would say, junior to mid product manager and even make backlog level decisions.

— Mikhail Shcheglov

I don't read my email anymore. Like an agent does it for me and pings me in case if it's, uh, if it's urgent. I don't manage my calendar anymore.

— Mikhail Shcheglov

summarization actually hurts a retrieval.

— Mikhail Shcheglov

I have good news for you. I think, uh, not only product management is going to exist, I think it's gonna thrive.

— Mikhail Shcheglov

Company knowledge graph and context digitizationProduct context coverage metric (industry/business/customer)Agentic CPO responsibilities and operating system mindsetSlack agent: status reports and feature request gatekeepingOpenClaw + Hermes architecture and skill generationThree-layer memory: graph, vectors, raw transcriptsImperatives/guardrails to prevent hallucinations and “fake helpful”Hybrid retrieval and token optimization/model routingBoard skill for strategy/pitch deck critiqueAI-generated design system maintenance loopBacklog abstraction across team spreadsheetsRecruiting automation with LinkedIn/CRM + transcript-based evaluationFuture of PM roles and hiring AI-native PMs

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

iOSAndroidClaudeChrome