Agentic Workflows
An agentic workflow layer over the S2 ontology: the agent roster from the 2026-06-11 fusion assessment, expressed as Arazzo shaped pipelines. Prewired and dry runnable today; wiring to an approved agent runtime or Foundry AIP is a one file swap.
An agentic workflow is a declared, multi step sequence an AI agent executes toward a goal: retrieve grounding data, reason over it, act through scoped operations, and pause for humans at named gates. The agent runs the loop plan, ground, execute, reflect, summarize. The point: agent capability lives in open, auditable definitions with scoped credentials, not in prompt spaghetti. This is the pattern Palantir AIP, AWS AgentCore, and Google ADK all converged on.
Arazzo (an OpenAPI Initiative standard) declares a workflow as inputs, ordered steps, and outputs. Each step binds to an operation, states its successCriteria, and names what it passes forward. Deterministic, portable, reviewable. Every definition on this page is that exact shape in TypeScript, with IntelClient methods standing in as the operations.
The 2026-06-11 agentic fusion assessment replaced the idea of one smart assistant with a roster of narrow specialists: RFI Triage, Collection Gap Watcher, Correlation, Indications and Warning, the Brief Drafter under a Supervisor, and the Lineage Recorder. Each workflow here is one roster row made executable: one job, one slice of the ontology, one human gate.
Every workflow is fully defined and dry runnable today: retrieve steps perform real reads against the live ontology, reason steps run deterministic heuristics that cite their records, and act steps are captured as simulated writes instead of performed. Wiring (an approved agent runtime, a scoped credential, and published OpenAPI plus Arazzo documents) flips one selector in src/lib/agentic/index.ts. No view changes.
Agents draft, correlate, watch, and recommend. Humans decide. Every workflow pauses at a named gate (Collection Manager, fusion analyst, cell lead, S2) before any write becomes real, and every claim a run makes traces to ontology records: the no invention contract. The commander approves judgments, never keystrokes.
Two destinations, same seam. Standalone: the registry's bindings become real Arazzo documents an approved runner executes with scoped credentials. MSS/Foundry: each workflow re homes as an Agent Studio agent or AIP Logic function, gates become approval tasks, acts become scoped ontology Actions, and lineage is native. The definitions survive both moves because they are data, not code.
An RFI arrives. The agent proposes the PIR it supports, the cell that should own it, and a draft EEI, all grounded in current holdings so duplicates are caught before they cost a collection asset.
Standing watch over the gap register and PIR coverage. Raises open gaps that have no collection request working them and drafts the request for the CM to validate.
Reads across all reporting and proposes that separate reports describe one actor, one place, or one activity, by shared entity, BE number, and grid. Confirming a link is a click; finding it by hand is the work that gets missed.
Stands watch on the indicator list from the running estimate and flags reporting that touches an indicator that has not yet been observed. Pushes to a human, never acts alone.
Checks a draft analyst report against the production template (DTG, grid, recommendation, EEI status, grounded summary) and stages the lifecycle move to Pending Review with a fix list when the template is short.
The flagship. The Supervisor gathers cell assessments, coverage, gaps, and decisions in parallel; the Brief Drafter seeds the BLUF and posture from live counts; the analyst edits; the S2 approves; the archive freezes the lineage.