I Gave an AI Agent the Keys to My Life (Here's What Happened) — Radek Sienkiewicz (@velvetshark-com)
Summary
Radek Sienkiewicz, an OpenClaw maintainer, presents a personal account of incrementally delegating control of his digital life to an AI agent running on OpenClaw — a personal AI agent framework — over the course of months. The talk is framed as a practical case study from the perspective of a power user and contributor rather than a systems architect.
The core thesis is that trust and capability should be extended to an AI agent one small, reversible step at a time. Sienkiewicz began with a single communication channel (WhatsApp, later migrated to Telegram, then Discord) and added capabilities only after each prior step proved stable. This incremental model meant he never experienced the catastrophic failures — bricked systems, broken updates — that he observes in others who adopt large configurations all at once. If something breaks, stepping back one increment, diagnosing the failure, and hardening against recurrence is the consistent pattern.
The most consequential expansion was integrating his Obsidian knowledge base, which contains approximately 3,000 markdown notes covering work projects, personal research, tasks, articles, and bookmarked links. When a link is added to an inbox note, the agent fetches the content (tweets, threads, articles, YouTube videos), adds tags and context, queries the vault for related existing notes, and surfaces connections the user had forgotten — essentially converting passive bookmarks into an actively maintained knowledge graph. Sienkiewicz notes this mirrors what Andrej Karpathy described in a viral tweet about LLM knowledge bases, and reflects that he arrived at the same architecture organically, step by step.
The agent also runs an overnight cron ecosystem (roughly 3–6 AM) that performs automated maintenance: indexing the vault, refreshing QMD (query-based memory) indexes, backing up memory so data loss is bounded to a few hours of work, and pulling the latest OpenClaw version using scripts that know what can break, how to verify a safe update, and how to restart the gateway reliably. The user wakes to a fully updated, indexed system with a morning briefing of email and calendar.
Sienkiewicz categorizes his agent's work into three job types: Ambient Operations (background plumbing and automation that requires no judgment), Attention Filtering (proactively surfacing urgent signals — a failed Netflix payment resolved in five minutes, a domain renewal that would have been missed, or email drafts pre-populated with project context drawn from Obsidian), and Execution Support (drafting, synthesizing, and consulting tasks organized by purpose-specific Discord channels including inbox, consulting, video research, briefing, Instagram, YouTube, OpenClaw maintainer work, and a rotating playground channel for experimentation).
On the architecture side, LLMs handle judgment-intensive tasks — understanding email context, making knowledge connections — while deterministic scripts handle fixed conditional logic without invoking an LLM at all. Memory management is a recurring challenge: Sienkiewicz uses a tiered system including a soul.md file, a critical-rules.md file (placed high in the agent configuration to prevent forgetting), a memory folder, and a dreaming mechanism for memory promotion. He warns that bad memory compounds at scale — thousands of notes without active curation become a liability — and that multi-step automations (10+ steps) are brittle and should be decomposed or guarded with explicit failure handling. Noisy nodes are cleaned regularly.
The talk closes with a philosophical framing around optimizing for the "future self": rather than treating tomorrow's version of oneself as a god-creature who will handle deferred tasks, the agent's role is to complete as much as possible on behalf of that future self overnight, so that each morning begins with maximum context and minimum overhead. Sienkiewicz encourages attendees to start with one recurring pain point, grow trust incrementally, move as much as possible into markdown files, and maintain an inspectable, editable system throughout.
Key Takeaways
- •Extend agent permissions one small step at a time — each new capability should be stable before adding the next, so failures are isolated and reversible.
- •Store all personal and professional context (projects, tasks, research, bookmarks) in markdown files so the agent can query, cross-reference, and surface connections automatically.
- •Use an overnight cron window (3–6 AM) for indexing, backups, and software updates so the agent is always current and data loss is bounded to a few hours.
- •Separate judgment tasks (handled by the LLM) from fixed conditional logic (handled by deterministic scripts) to avoid unnecessary model calls and improve reliability.
- •Place critical behavioral rules in a dedicated critical-rules.md file and reference it high in the agent configuration to prevent the agent from forgetting hard-won constraints.
- •Decompose automations beyond ~10 steps into simpler sub-automations or add explicit guardrails, because long chains will eventually break.
- •Actively curate memory files — at thousands of notes, uncurated memory compounds into a performance and accuracy liability.
- •Use a rotating playground Discord channel to safely experiment with new models, workspaces, and memory configurations without risking production channels.
- •Frame the agent's purpose as optimizing for the future self: automate everything that can be done in advance so each day starts with maximum context already prepared.
- •An inspectable, file-based system (markdown, editable scripts) is easier to debug, tune, and trust than opaque black-box configurations.
Transcript
[music] >> Ready to go? All good, yeah. Please welcome Radek, everybody. >> [applause] >> Hey. I'm Radek and one of the Open Claw maintainers. And uh I want to talk what happens like in my life with Open Claw when I practically gave the keys to my life to to Open Claw. Uh and like it's almost like literally. And uh what that actually means. So so this happens like step by step. It wasn't all at the same time, but it can access my emails, it can access my notes, files, calendars, tools, my operat...
Topics
Related Sessions
