Aakash GuptaHow to build a Company Operating System with Hermes and OpenClaw
At a glance
WHAT IT’S REALLY ABOUT
Building an agentic company operating system with knowledge graphs and skills
- OLX’s CPO built an “agentic Company Operating System” integrated with Slack, Google Workspace, Confluence, Jira, and meeting transcripts so stakeholders can query projects, validate feature ideas, and even prototype work through an AI gatekeeper.
- The core asset is a continuously growing company knowledge graph and three-layer memory (graph + vector DB + raw transcripts) designed to prevent knowledge leakage and enable higher AI autonomy as context coverage increases.
- They manage the COS like a product: tracking “product context coverage,” running recall/accuracy evaluations, and iterating quickly on rules, tools, and model routing to balance quality and token cost.
- Operational work (status reports, calendar/email triage, intake, recruiting workflows) is delegated to agents so PMs spend more time on customer discovery and value discovery rather than “process theater.”
- Mikhail argues PM roles won’t disappear but will shift toward discovery and orchestration, with smaller teams and blurrier lines between PM/engineering/design as all become quality-and-token-budget managers.
IDEAS WORTH REMEMBERING
5 ideasContext is the new company asset; prevent knowledge leakage by making it machine-retrievable.
Mikhail frames the goal as “digitizing” business/product/customer/technical context into an always-available system so it doesn’t walk out the door when a key person leaves. More context in memory enables higher autonomy: the agent can move from clerical work to backlog- and eventually strategy-level help.
Measure your Company Operating System with a context-coverage KPI.
They track “product context coverage” (e.g., 54%) by prompting the agent to estimate how much it understands across industry verticals, business model/PNL drivers, and customer segments. While imperfect, it becomes a directional KPI that rises as PMs load more research, RFDs, and transcripts.
Three-layer memory beats summaries; raw transcripts improve recall.
The COS uses (1) a knowledge graph, (2) a vector DB for fuzzy retrieval, and (3) raw transcript archives of every conversation/meeting. Counterintuitively, they found summarization reduces retrieval fidelity (lost nuance + forced templates), so they store raw transcripts because storage is cheap.
Hybrid retrieval (keyword + vector) keeps answers accurate and token-efficient.
They use hybrid retrieval: try keyword match first, then vector search for ambiguous queries, pulling only the most relevant snippets to avoid bloating the context window. This is how the agent answers questions like project status without ingesting entire meeting histories.
Reliability comes from strong behavioral guardrails, not just better models.
A large “imperatives” library (hundreds of lines) constrains behavior: no fabrication, think-before-act, facts-over-guesswork, avoid “fake helpful” responses, etc. Priority is handled via files like CLAUDE.md (highest priority, kept short) and SOUL.md (longer, OpenClaw-specific).
WORDS WORTH SAVING
5 quotesAgents are actually building most of our features. Any stakeholder can talk to an agent and validate any feature request they want.
— Mikhail Shcheglov
The core thing that we're trying to prevent here is the leakage of that knowledge.
— Mikhail Shcheglov
Summarization actually hurts a retrieval.
— 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
Not only product management is going to exist, I think it's gonna thrive.
— Mikhail Shcheglov
High quality AI-generated summary created from speaker-labeled transcript.
