Published on

AetherCrew: An Operating Paradigm for AI-Native Startup Execution

Authors
  • Name
    Motions Technologies
    Twitter

AetherCrew: Operating Paradigm

AetherCrew is not a chatbot bolted onto a backlog. It is an internal AI startup team — named agents, durable state, and two hard human gates — designed so a founder can ship production software with research-grade control loops.

The loop

TriggerBriefGO?ExecuteVerifyShip or DraftDigest

Six states. Two humans-in-the-loop: GO on a brief, and SEND on any outbound blast-radius action. Everything else is agents talking to agents on a durable graph.

StateOwnerCost signal
TriggerIntakeFree / ms
BriefIris (Product)Cents of LLM
GOHuman$0 while waiting (Dynamo checkpoint)
ExecuteAtlas plans; Kai/Ren codeSpot minutes + tokens
VerifyPriya (QA), Sig (Sec)Cheap LLM + tests
DigestIrisTimeline + postmortem

If a state fails, the graph checkpoints, pages the right owner, and waits. Nothing loops silently. Nothing retries a paid model without FinOps approval.

What “agent” means here

An agent is:

  • A name Jay can address (“Atlas, why Postgres?”)
  • A role and a never-do list — Atlas plans, Atlas does not commit; Nora drafts, Nora does not send
  • One brain class — Grok 4.6 High on decision nodes; roles differ by prompt and tools
  • A subgraph — LangGraph node with interrupt points, not a 24/7 process

An agent is not an always-on Python service. That is a bill without a purpose.

Team = supervisor graph

The team is a supervisor graph, not a group chat. One graph routes work; each agent returns a command or message; the supervisor picks the next hop or interrupts for the human.

Swarm-style peer handoff is reserved for the on-call subgraph (SRE ↔ Security) where round-tripping through the supervisor would add latency without safety.

Scientific stance

Traditional sprint theaterAetherCrew
Ticket sits in To Do for a weekBriefed in tens of seconds, then GO or wait
Estimation is a meetingToken + minute forecast on the brief
Standup finds blockers tomorrowGraph pages the owner when a node fails
Done = mergedDone = merged or drafted-and-linked with reason
Retro every two weeksEvery ticket writes a one-paragraph postmortem

This is systems research applied to execution: state machines, budgets, provenance, and falsifiable failure modes — model outage, runaway recursion, stuck sandbox, wrong brief, wrong draft PR.

What we are trying to achieve

  1. Production leverage — ship UmaMeats / HandyGo / Vault-class systems with an AI team that respects money and blast radius.
  2. Auditable autonomy — after GO, agents decide, retry, and disagree in the trace; interrupt only on red lines.
  3. Meeting-native collaboration — the same roster can stand in a virtual room with factored speech (see UM-SMF), without a second orchestrator.
  4. Continuity — Mneme (world model + compile + initiative) so the team does not reset to amnesia every turn.

AetherCrew is Motions’ bet that the next competitive edge in software companies is not more seats in Jira — it is a measurable agent operating system.