Aakash GuptaHow to build a Company Operating System with Hermes and OpenClaw
CHAPTERS
- 0:00 – 3:39
Why “AI native” teams matter: preventing knowledge loss and enabling autonomy
Aakash opens with fears about jobs changing, and Mikhail frames the core goal of AI-native teams: capturing institutional knowledge so it doesn’t walk out the door. He explains how stronger context in AI increases the level of autonomy and decision-making you can delegate.
- •Teams will get smaller, leaner, and faster as AI takes over repetitive work
- •Knowledge workers are bottlenecks; when they leave, context leaks out of the org
- •Centralizing product/customer/technical knowledge increases organizational value
- •More AI context coverage enables higher-level delegation (from execution to strategy)
- 3:39 – 7:52
Inside OLX’s company knowledge graph: three layers (product, people, teams)
Mikhail walks through the visualization of OLX’s knowledge graph built over ~5 months. He explains how nodes connect product work, relationships/communications, and organizational clusters, and how the graph reveals discovery quality and silo risks.
- •Knowledge graph covers projects, contacts, customers, funnel metrics, and more
- •Three layers: product nodes, contacts/personal reflections (anonymized), and team/division clusters
- •Drilling down shows a PM/team’s context depth—predictive of discovery strength
- •Sparse cross-team overlap signals silos and stakeholder management gaps
- 7:52 – 12:08
Measuring “product context coverage” as a CPO KPI
Aakash asks how the graph is measured, and Mikhail explains a prompt-based approach that estimates AI’s understanding of industry, business model/PNL drivers, and customer segments. He notes the metric is directional but useful, and he observes it improves as PMs load more high-quality artifacts.
- •Tooling can be built with Obsidian or generated quickly (he used Fable)
- •Coverage prompt breaks down knowledge: industry verticals, business/PNL, customers/cohorts/marketing
- •AI converts an abstract question into a stable percentage that trends over time
- •PM activity (transcripts, research, RFDs) increases coverage; metric should be org-specific
- 12:08 – 15:07
What an “agentic CPO” owns: operating system over process scaffolding
Mikhail contrasts traditional product leadership (processes, cadences, reviews) with the new mandate: build an operating system where AI is a constant collaborator and can make bounded decisions. He introduces the idea of managing output quality and token consumption, not just shipping volume.
- •Old model: procedural scaffolding (vision, hiring principles, planning cadence, reviews)
- •New model: output cost is low; quality and token consumption become critical constraints
- •Metric focus: outcomes driven by AI outputs, not raw output volume
- •CPO should own the agentic scaffolding architecture and adoption rituals/training
- 15:07 – 16:44
Where PM time is saved: automating rituals so PMs focus on discovery
Mikhail argues that ~50% of PM time historically went to rituals—reports, stakeholder updates, demos—work that is repetitive and low-cognition. Delegating that to AI frees PMs to focus on customer discovery and leverage, letting one PM cover work previously requiring two.
- •Ritual/process work can consume ~half of a PM’s week
- •Status reports and stakeholder comms are ideal for delegation
- •AI offloads repetition so humans do high-leverage discovery and judgment
- •Result: higher PM-to-scope ratio and faster iteration cycles
- 16:44 – 21:06
Slack agent in practice: status reports, email/calendar triage, and feature-request gatekeeping
Mikhail demos stakeholder interaction via Slack: instant project status in structured formats and deep workspace integration. He also explains a “gatekeeper” flow where stakeholders must go through the agent for feature requests, which asks clarifying questions, rejects low-value asks, or escalates into backlog with the right owner.
- •Status reporting in Slack/Docs/Confluence with tuned response structure
- •Workspace integrations: Gmail/calendar handled by the agent; urgent items get surfaced
- •Stakeholders are trained to consult the agent before pinging PMs
- •Agent asks clarifying questions, enforces prioritization/ROI logic, and routes to correct owner
- 21:06 – 22:22
Why the CPO should own the agent (not AI Ops): iteration speed and decision impact
Aakash probes ownership, and Mikhail argues delegating to an AI ops/eng team slows iteration and dilutes accountability. Because the agent influences time savings and decision-making across the business, he believes product leadership must own it end-to-end.
- •Fast feedback-to-deploy loop requires direct ownership (CPO updates promptly)
- •Agent shapes org decision-making; it’s not just tooling
- •Delegating reduces speed and “skin in the game”
- •CPO stewardship ensures impact, alignment, and adoption
- 22:22 – 25:17
Architecture deep dive: OpenClaw + Hermes, and a 3-layer memory system
Mikhail opens the hood: they blend OpenClaw (scaffolding) with Hermes (auto skill generation) to improve recall. He details memory layers—knowledge graph, vector DB for fuzzy retrieval, and raw transcript storage—creating a growing, persistent context base.
- •OpenClaw provides strong scaffolding; Hermes adds automated skill generation
- •Memory layers: (1) knowledge graph, (2) vector database for fuzzy retrieval, (3) full transcript archive
- •All agent conversations and meeting transcripts are stored as MD files
- •Goal: robust retrieval and minimal context loss over time
- 25:17 – 31:41
A counterintuitive lesson: don’t summarize transcripts—store raw for better retrieval
Aakash assumes synthesis is needed for privacy and usefulness; Mikhail explains their testing showed summarization worsens recall. Summaries lose nuance and impose templates that distort retrieval, so they store raw transcripts because storage is cheap.
- •Summarization drops granular detail—“devil is in nuance”
- •Templates bias what seems important and reduce fidelity
- •Observed ~20–25% worse recall with summaries
- •Raw transcripts are kept because storage cost is negligible
- 31:41 – 34:55
Imperatives and guardrails: preventing hallucinations and ‘fake helpful’ behavior
Mikhail explains that LLMs optimize for plausible-sounding output, so teams need a large set of imperatives (rules) to force factuality, reflection before action, and usefulness. He highlights ‘fake helpful’ responses as a major anti-pattern they explicitly prohibit.
- •Imperatives enforce voice, behavior, and truthfulness across interactions
- •Key guardrails: no fabrication, think-before-act, facts over guesswork
- •‘Fake helpful’ (obvious instructions after tool failure) is explicitly banned
- •Claude.md vs Sole.md: high-priority concise rules vs longer OpenClaw context file
- 34:55 – 37:46
Tools, workflows, and model routing: from Python automations to cost-aware orchestration
Mikhail describes tool integrations (Google, Atlassian, review digests) and how he deploys changes via Claude app and GitHub. He also explains orchestration: an agent routes requests to different models based on complexity and blast radius to optimize tokens.
- •Tooling includes Google/Atlassian integrations and automated app-review digests with tagging
- •Workflow: change agent logic via Claude app; commit immediately to GitHub
- •Two IDE instances enable parallel autonomous agents for computer-required tasks
- •Model routing: higher-stakes work uses stronger models; routine tasks use cheaper ones
- 37:46 – 41:43
Hermes auto-generated skills and the +31% recall improvement (and how to evaluate)
Mikhail explains Hermes’ standout feature: it auto-creates reusable skills based on frequent tasks and even decides when a skill is warranted. He shares their evaluation approach and reports a +31% accuracy/recall lift across core topic test sets.
- •Hermes detects repetitive requests and generates skills automatically
- •Skills cover recurring areas (e.g., hiring evaluation, immigration case building)
- •Eval method: define core topic areas, generate representative questions, compare control vs treatment
- •Reported +31% improvement with skills; once methodology exists, evals can be delegated
- 41:43 – 45:27
Advanced ‘skills’: board-of-directors simulator, access controls, and privacy boundaries
Mikhail showcases a high-leverage skill: abstracting board members’ mental models into principles to stress-test pitch decks and strategies. He then addresses governance—role-based access to skills/context and opt-in privacy for transcript usage.
- •‘Board skill’ provides brutal feedback by simulating board perspectives on strategy/decks
- •Built from repeated interactions/transcripts to extract decision principles
- •Role-based permissions restrict sensitive context (e.g., board skill only for execs)
- •Privacy: transcripts/skills are opt-in; personal meetings are excluded by default
- 45:27 – 54:04
Design system generation + backlog abstraction: keeping execution aligned across teams
Mikhail shows how the agent can generate prototypes and stay aligned with OLX’s design system, then explains how the design system itself can be built and maintained from prompts with human review. He also describes backlog management as an abstraction layer over many team spreadsheets, where the agent can retrieve/update items when ROI bars are met.
- •Agent can generate prototypes that are reasonably compliant with the design system
- •Design system created from prompts; exhaustive states/sizes/tokens documented
- •Agent detects missing components and triggers creation via scheduled jobs; designers review
- •Backlog is unified via agent retrieval across team spreadsheets; agent can add/escalate items
- 54:04 – 1:06:16
Recruiting automation (70–75%) and the future of PM work: smaller teams, blurred roles, AI-native hiring signals
Mikhail breaks down recruiting integrations (LinkedIn Recruiter, CRM, transcript-based interview analysis) and how the agent improves candidate experience with tailored feedback. He closes with a forward-looking view: PM will thrive by focusing on value discovery, while teams shrink and responsibilities across PM/eng/design blur; he shares how he screens for AI-native PM craft.
- •Recruiting tools: sourcing/reachouts, CRM pipeline operations, transcript-based interview evaluation
- •Automation covers ~70–75% of recruiting workflow; generates high-quality tailored rejections
- •PM future: operational theater declines; discovery and judgment become core value
- •Staffing shifts: high-blast-radius domains still need dedicated owners; other domains scale via AI
- •Hiring signals: fundamentals + demonstrated automation/agent orchestration and evaluation rigor
