Le Nain family, after Le Maître des Jeux, A Family Meal, c. 1645–55 or later. The Cleveland Museum of Art, 1958.174. CC0.
Host owns the menu.
Good morning.
First Friday on the clock since the return letter. This week’s bookmarks stopped pitching another mega-agent and argued for control surfaces: who prepares the options, what refuses silent corruption, how agents call each other without a public webhook, and which cloud path actually returns attention.
Host owns the menu. Everything else is costume.
Host owns the menu.
Avid wired Jev as a typed decision layer under Opus 5.5 and claims roughly 80% less cost and time in their coding harness. Opus does the hard reasoning. Jev picks from options the host prepared and the harness validates — which notes to load, which worker gets the task, which recovery path when a tool fails, which checks before the full suite. Named options. Checkable results. Logged outcomes. The small model cannot invent providers or tools. “Answer” focus gets an empty tool bundle.
Same pattern as the cheap-router keepers earlier in the month. Not a new religion. Put an inspectable picker in front of the expensive model. Keep the evidence.
Harness software beats prompt skills.
Paper Instruments open-sourced Paper Office: agent-first forks of the usual Python Office libraries. Familiar imports. Typed search and spans. Package-preserving saves. Guarded edits. Refusals instead of silent OOXML corruption. On 61 brownfield tasks across five models, Paper plus guidance passed 92.5% of trials versus 80.7% upstream and 69.5% with Anthropic’s Office skills. Agents fell back to raw OOXML editing in 1.6% of Paper runs versus most of the time without it. Expose the hidden structure. Validate before you mutate. Preserve the parts you did not mean to touch.
Hedgineer makes the complementary call for generative widgets: Claude Skills when the shape must rewrite a known template around live user data; fixed MCP Apps when the surface must stay deterministic. Use Skills when the data is messy. Keep MCP Apps when inventing a shape would be a compliance bug.
Private buses. Cloud that buys back attention.
Matt Van Horn got one agent to ask another to place a phone call over Tailscale, not a public webhook. Heterogeneous agents requesting work under the user’s network identity. The missing piece is a private bus, not another mega-agent brand.
Beeper bot-teams is the chat cousin — profiles and shared channels instead of one bot in a DM.
On the attention side: Amp Orbs as managed machines that boot with the agent runtime and project so laptop babysitting stops eating prompts. Zuse as an open workspace that runs the agents you already pay for, with sandboxed git worktrees. DigitalOcean Managed Agents as Firecracker microVMs with pause/resume and a governed tool gateway. Shared question: does the cloud path remove lid-close drama, or invent a second harness? Pause economics and env parity matter more than the logo on the box.
Guilherme Castro · swaraj · DigitalOcean
Try this
Three moves. Own the menu.
- Pick one real decision surface. Write the menu first — named options, validators, logged outcomes. Put a cheap typed picker in front of the expensive model. Measure cost and time.
- Anywhere agents touch Word, Excel, or PowerPoint: trial a harness that refuses silent corruption before another prompt skill. Prefer refuse over quiet damage.
- Wire one agent-to-agent hop on a private mesh. Compare cloud runtimes only on attention returned and pause economics — not brand.
Host owns the menu. Keep the gates. Leave the costume on the shelf.