Continuations

What a continuation is

Every framework today treats model.invoke() as the atomic unit of work. Graphs wire those calls together; orchestrators dispatch between them. The problem is that this makes the agent terminate: a request comes in, work happens, a response goes out. Research doesn't work that way. A nutrition consultant evaluating a feed additive over six weeks needs something that continues.

A continuation is the atomic primitive that lets that happen. It carries:

• Identity — stable id, lineage, generation
• Goal frame — a structured representation of what it's trying to accomplish (not a prompt)
• State — working draft, open questions, accumulated findings
• Budget vector — tokens by tier, dollars, wall-clock, tool quotas, human-attention quotas
• Wake conditions — predicates that say when to schedule it next
• Context binding — pointer to its field projection
• Lineage edges — parent, children, siblings, merge targets

The hot store schema

Cosmos DB on Azure, DynamoDB on AWS — same logical model behind a substrate interface.

continuation
  continuation_id          UUID, PK
  root_id                  UUID
  parent_id                UUID NULL
  generation               INT
  tenant_id                UUID
  researcher_id            UUID
  status                   ENUM
  created_at               TIMESTAMP
  updated_at               TIMESTAMP
  next_wake_at             TIMESTAMP NULL
  goal_frame               JSONB
  state                    JSONB
  budget                   JSONB
  wake_conditions          JSONB
  context_field_id         UUID
  policy_class             TEXT
  tags                     TEXT[]

continuation_event           # append-only event log
  event_id                 UUID, PK
  continuation_id          UUID, FK
  ts                       TIMESTAMP
  kind                     ENUM
  payload                  JSONB
  cost                     JSONB

The goal frame is structured

This is the contract between the researcher and Maestro. It's also what fingerprints the continuation for the context field — semantically similar frames share context.

{
  "intent": "evaluate_feed_additive_effect",
  "subject": {
    "species": "dairy_cow",
    "lactation_stage": "early",
    "cohort_size_range": [60, 200]
  },
  "question": "Does {additive} at {dose_range} improve {outcome} vs control?",
  "bindings": {
    "additive": "monensin",
    "dose_range_mg_per_day": [200, 400],
    "outcome": "milk_yield_kg_per_day"
  },
  "evidence_requirements": {
    "study_types": ["rct", "field_trial"],
    "min_n_per_arm": 30,
    "min_recency_years": 10
  },
  "deliverable": "structured_evidence_brief",
  "success_predicate": "evidence_brief.confidence >= 0.7 AND human_signoff.received"
}

Wake conditions are why this works

A sleeping continuation specifies why it should be woken. A continuation can sleep for three weeks waiting on an arXiv RSS event, wake for five seconds of work, and go back to sleep — without anyone "running" it.

{
  "any_of": [
    {"kind": "timer", "at": "2026-05-20T09:00:00Z"},
    {"kind": "event", "stream": "arxiv.q-bio", "predicate": "matches_goal_frame"},
    {"kind": "human_signal", "from": "researcher_id", "topic": "approval"},
    {"kind": "sibling_publish", "root_id": "...", "tag": "finding"},
    {"kind": "data_arrival", "source": "cargill.feed_trial_2026_q2"}
  ]
}

Migration and failure semantics

Continuations are designed to be migratable across workers. Cooperative locking with fencing tokens, idempotent event log, no migration mid-model-call — model calls are checkpoint boundaries. The Erlang and Temporal lineage, adapted for LLM-driven work.

Why this matters for Cargill

A nutrition consultant submits one goal frame on Monday. The root spawns eleven children: three pulling structured trial data, four monitoring literature streams that will fire for weeks as new papers land, two synthesizing evidence with Opus-class reasoning, one drafting the regulatory section, one watching a competitor's RSS feed. Eight go to sleep within ten minutes. By Friday, two have woken because new trials were published. Total spend: ~$14. Total researcher time: ~35 minutes, mostly review. That's the work continuations enable.