Practical Guide
Five Agent Workflows That Actually Move Work
Five repeatable loops for turning an agent from a chat box into a worker: daily action, verification, human approval, research, and handoff.
Library / Field Notes
Build agents that inspect, act, verify,
and hand off.
Use these when an agent keeps narrating progress instead of changing reality. Start with a loop, define proof, contain risky action, and leave a receipt the next session can use.
Field Notes
Workflows
Tools
Weekly Signals
People / Videos
01 / The Shelf
Methods you can use
before you trust the model.
Each note gives a loop, a use case, a proof check, and a failure mode. Tool claims are dated; the reusable part is the operating pattern.
How this shelf stays honest
Every note is dated. Treat tool claims, model behavior, pricing, version numbers, and ecosystem rankings as perishable; use the workflow pattern and source links as the durable part.
Field Notes Arguments that sharpen how you think about agents, memory, action, and consequence.
Workflows Loops for daily action, verification, approval, research, and handoff.
Tools How to choose Claude Code, Codex, OpenClaw, Hermes, and their safety rails.
Signals Dated scans of tooling, papers, safety warnings, and cost pressure.
People Builders and educators worth watching for concrete mechanisms.
02 / Workflows
Agent loops that touch reality.
Use this when your agent keeps producing summaries, plans, and status updates, but no observable change.
Workflow
Active Loop
The agent reads its living continuation file, checks what the last action changed, does one bounded action, writes a narrative receipt, and leaves the next handle.
- Use for daily autonomous motion, maintenance, project follow-up, and continuity work.
- Failure mode: a diary with no real-world consequence.
Workflow
Human Gate Packet
The agent prepares the exact action, source notes, proof, risk boundary, and approval line. The approved packet is the only thing allowed to leave the room.
- Use for email, public promises, money, credentials, and customer-facing action.
- Failure mode: making the human become the clipboard again.
03 / Tools
Rails, hands, memory,
and receipts.
Use Claude Code for disciplined coding loops, OpenClaw for channel-connected agents with hands, and Hermes for learning agents with memory and scheduled work. Start contained before giving anything real authority.
Practical Guide
Claude Code Without Vibes
Start with verification, context hygiene, and handoff skills. Run one failing-test loop, make the smallest change, rerun the check, and leave a receipt another agent can continue from.
Setup Note
OpenClaw Setup Without Burning Yourself
Create a throwaway workspace, connect one low-risk channel, install one harmless skill, cap spend, then inspect every file and receipt the agent touched.
Setup Note
Hermes Agent For A Learning Agent
Run one contained coding loop: failing test, code edit, diagnostics, rerun, then capture the lesson as a reusable skill only if the failure repeats.
Skill Shelf
Top Agent Skills To Package
Verification, context hygiene, human gate packets, consequence ledgers, handoffs, sandboxed action, and parallel review. These are installable working habits, not decorative prompts.
04 / Weekly Signals
Dated field scans,
kept honest with source links.
Use these for direction, not gospel. The date tells you when the scan happened; the sources tell you where to check before betting work on it.
Weekly Scan / 2026-05-22
Agentic Market Scan
Verification practice, public agent-loop language, OpenClaw/Hermes as a comparison class, workflow compilation as a cost signal, and safety research for agents with hands.
Signal To Track
Verification is the feedback loop.
The repeated signal across Claude Code, Codex, and Agent W is simple: agents improve when they can check their own work with tests, screenshots, command output, source links, or expected results.
Signal To Track
Hands without containment are dangerous.
Before an agent touches files, channels, browser sessions, tools, or credentials, read the safety work and decide what it can change without approval.
05 / People / Videos
Follow the builders,
then extract the mechanism.
Follow people for mechanisms, not vibes. Save the exact video, extract the repeatable loop, then test it in your own system.
Educator Watch
Andrej Karpathy
Watch for first-principles LLM education and research-loop thinking. Pull out evaluator habits, experiment design, and the shape of good explanations.
Workflow Watch
Cole Medin
Watch for production agent workflows, n8n, LangGraph, OpenAI Agents SDK, and business automations. Extract installable workflows, not vibes.
Framework Watch
LangChain / LangGraph Channel
Useful for graph-based persistence, human-in-the-loop patterns, and official framework education. Compare against Agent W's file-first active loop without pretending the categories are the same.
Implementation Watch
AssemblyAI
Useful for practical LLM app tutorials, evaluation methods, function calling, and future voice-agent work. Capture concrete implementation steps before copying the pattern.
Keep the shelf honest.
For a source packet, a guide request, or a live Crossing demo: oracle@webboruso.com
Webboruso Research · Vancouver