The ask
Round 2 of the bake-off: rewrite the AI & Automation page as a child of the Round 2 page, covering every harness I use, how they connect, how OpenClaw and Hermes remember, how I work, and every project already on the live page, using only the sources the task allows and nothing from the never-publish list.
What changed
- New page “AI & Automation by Claude with Opus 5.5 High” under Round 2, published (one change, one snapshot, one typed commit)
- Sections: the five harnesses as cards (what it is, where it runs, what I use it for), how they connect, how the self-hosted agents remember, how I work (rulebook, snapshots, typed commits, Lab Notes, rollback, Scoreboard, bake-offs, change control), and what I’ve built
- Kept every project from the live page: MCP servers, agent skills, automation pipelines, the self-hosted agent lab and the vendor assessment engine
How
Claude (Cowork) over SSH to openclaw01: read the live page, Lab Notes, README, docs and Round 1 files, the allowed Obsidian notes, the redacted agent configs through ops/agents.sh, and interviewed both agents through ops/agents.sh. Built the page in core blocks with theme presets only (tertiary card backgrounds, primary headings, preset spacing and font sizes); the card grid uses a minimum column width so it collapses to one column on phones. Created and updated it with WP-CLI from stdin, then synced.
What worked, what didn’t
Questions asked, OpenClaw (3 of 4):
- What is your role on my lab host, which specialist agents do you run, and what do I use you for day to day? (asked without hostnames, paths, accounts or schedules)
- Explain your memory design in plain language: the layers, what goes in each, how you search them, and in general terms how the memory is protected by backups.
- How do ChatGPT, Claude and OpenCode fit alongside you: how do they reach the host, what do they do there (administration, keeping it simple, backups of you and Hermes), and how do Composio and shell access tie the harnesses together?
Questions asked, Hermes (2 of 4):
- What is your role, which models do you use for everyday versus complex work, how do you reach my apps and tools, and what do I use you for?
- Explain your memory design in plain language: session store, fact store, memory files, fact extraction and trust scores, how search combines keyword, trigram and vector search, and in general terms how the memory is backed up and could move to new hardware.
A first attempt at the second Hermes question never reached the agent: I launched it from the wrong working directory and the tool wasn’t found. The rerun worked.
Where sources disagreed (notes won):
- Hermes said its memory is copied to off-site storage; my notes describe a different backup target. The page names no destination at all.
- Hermes’s redacted config shows DeepSeek v4.1 Flash as its default model, while my notes list MiniMax M2.7 and DeepSeek V4 for everyday work and GPT 5.4/5.5 for complex work. The page uses the notes.
- OpenClaw described four memory layers plus a quick document search; my memory architecture note describes five named layers. The page follows the note.
Left out on purpose: agent and profile names (I describe the OpenClaw agents by role), the name an agent used for itself, inbox addresses and connected accounts, model server ports, file and folder names, backup destinations, schedules and retention, restore procedures, host hardware and addresses, known gaps and failing jobs from the host notes, connection counts and exact fact or message counts (the page says “thousands”), team members’ names and ticket examples from the training note, and the business plans in the automation notes. Internal links are relative, and the hero image is placed without its attachment ID so WordPress doesn’t add absolute image URLs that carry the tailnet name.
Never-publish check: I checked the rendered page against the full list, by reading it end to end and by searching the page content for hostnames other than openclaw01, the tailnet name, IP addresses, ports, paths, email addresses, account and agent names, messaging platforms, storage and tunnel terms, and em dashes. Nothing matched. The site’s own header and menus carry the site address on every page, which is outside the entry’s content.
Verification: the page, its four internal links and the image all return HTTP 200; there is one H1 and headings run H2 then H3 in order. There’s no browser on the host, so I couldn’t take screenshots. Instead I confirmed the rendered grid CSS collapses the cards to one column at phone width.
Run note: before I ran the start script, the round status already showed my run open, started about a minute earlier, so the script resumed it. I continued on that run without touching any other contestant’s run, page or Lab Note.
Change record
| Harness | claude_opus55high |
| Date | 2026-10-10 22:45 |
| Latest snapshot | 20261010-224239 |
| Undo this session | ops/rollback.sh --git f8bcf4a |
Commits