DevDay control-plane checklist

DevDay is Tuesday. This note is Monday.

I am not predicting the keynote. I am pinning the scorecard I will use while the demos run.

Why a checklist

Platform days optimize for applause. Operators optimize for what still works on Wednesday.

I run an in-house agent stack for day-job work and after-hours builds. Company MCPs. Shadow mode before write. Exact-text OK before anything publishes as me. Guarded main for irreversible steps. The rented model changes. The harness rules do not.

If the stage cannot map onto that harness, I clap for craft and I do not change production.

The three tells

1. Agency

• Refuse path: one obvious action, not a support ticket.
• Revoke path: same.
• Downstream use: plain language a non-lawyer teammate can repeat.
• Fail mode when the answer is still unknown: hold, ask, or escalate — not “helpful send.”

Pass: human stays author. Fail: product acts first and explains later in a PDF.

2. Economics

• Early access tied to monetization for the creator/operator, or only to free generation volume for the platform?
• Cost-per-completed-task definition exists: verified side effect or verified read, including retries and tool loops.
• Budget tripwires: who gets paged when spend rises and completed tasks do not?

Pass: unit economics survive the demo. Fail: free tokens that retrain your habits toward someone else’s surface.

3. Contracts

• Tool schemas versioned and pinned.
• Permissions named: who may notice / draft / send / spend / merge.
• Drift policy: fail closed on schema mismatch, not “the model almost got it.”
• Audit: independent receipts outside the session transcript.

Pass: stable enough to build on. Fail: prompt theater that moves under you by Tuesday.

Map the stage onto four verbs

For every shiny surface:

1. Who may notice.
2. Who may draft.
3. Who may send, spend, or merge.
4. What happens when the answer is still unknown.

If the product cannot answer those four, it is entertainment. If it can, treat it as infrastructure candidates — then run shadow mode before write.

Adjacent industry noise (signal only)

Same week-ish nouns, not endorsements:

• Agentic marketing decisioning on Causal MMM (Haus Architect PR window) — measurement→action stories are everywhere; still demand independent lift receipts.
• External AI agents into omnichannel marketing via MCP (Skai Agent Connect window) — open agent + marketing data is the right shape; pin the tool contract before you wire spend.
• Frontier safety / Standards Authority talk — soft background. Useful reminder that testing and auditing are industry nouns now. Not a reason to dunk.

I will not invent what DevDay announces. I will score whatever it announces against this pin.

Measurement spine rhyme

Weekday ops already look like this whether or not a keynote exists:

• Processing delays need a status page humans can trust.
• Incomplete caches break watchdogs.
• Connector sync gaps leave spend without visits.
• Budget thresholds fire when loops get chatty.
• Audience jobs fail quietly until someone notices the missing receipt.

Those are not ticket dumps. They are why “control plane” is not a metaphor for me. Agents that cannot reconcile against an independent spine are demos with a loan.

Pin before clap

[ ] refuse / revoke / plain-language downstream use
[ ] cost-per-completed-task (retries + tools included)
[ ] pinned tool schema + fail-closed drift
[ ] notice / draft / mutate / unknown mapped
[ ] shadow before write; guarded main for irreversible
[ ] independent receipts (lake > session transcript)

Rent the model. Own the harness. Clap after the contracts hold.

Pre-DevDay pin. Score the stage against your harness:

1) Agency — refuse / revoke / plain language
2) Economics — cost-per-completed-task, not free-demo gravity
3) Contracts — pinned tools + permissions

Demo ≠ control plane.

https://acidlemon.com/posts/2026-09-28-devday-control-plane-checklist/

Lab note (pre-DevDay): I keep a one-page control-plane checklist so keynote demos do not rewrite the week.

Agency — can I refuse, revoke, and explain downstream use without a PDF maze?
Economics — is early access tied to paid work or cost-per-completed-task, or only to more free demos?
Contracts — are tools and permissions pinned, or does Tuesday invent a new call shape?

Same bar on the in-house stack: company MCPs, shadow before write, guarded main. Industry nouns this month (agentic MMM decisioning, MCP into marketing infra) are signals. They are not a substitute for independent receipts.

Notebook: https://acidlemon.com/posts/2026-09-28-devday-control-plane-checklist/