UNCLASSIFIED
UNCLASSIFIED
ISS LLC · SDVOSB (cert in progress) · CAGE 9VKK3 · UEI C7YDV3P8EHL7 · Who we are · Watcher health · SHIELD/ATLAS home

SHIELD/ATLAS — Agentic Flow & Autonomy

This page is the direct answer to one specific line in warfighter SME feedback on the CAPES brief: "3 of CAPES Brief can be solved with APIs built within MSS leveraging AIPs and foundry to create agentic AI workflow."

Position. the warfighter SME is correct. MSS, AIP, and Foundry can orchestrate an agentic workflow on their own. They are workflow engines and they are excellent at it. What they do not ship with is the workflow content — the agents that already know what to collect, when to collect it, how to enforce citations, where to publish, and what the human-on-the-loop checkpoint looks like. SHIELD/ATLAS is that content layer. Of the 20 production agents in the catalog, 20 run autonomously on cycle today. They expose OpenAPI tool specs that AIP picks up natively, Foundry actions that Foundry ontologies inherit, and CoT events that TAK / JBC-P render on the operator's existing map.

The autonomy loop — what runs without an operator

Every ATLAS agent follows the same six-stage loop. The loop is identical to OODA with an explicit record stage for DoD AI governance and a learn stage that closes the hallucination feedback loop.

                          +-------------------------+
                          |   1. OBSERVE            |
                          |   collectors poll       |
                          |   external feeds on     |
                          |   their own cycle       |
                          +-----------+-------------+
                                      |
                                      v
                          +-------------------------+
                          |   2. ORIENT             |
                          |   source-provenance     |
                          |   classifier scores     |
                          |   reliability + credibility (NATO STANAG 2511)
                          +-----------+-------------+
                                      |
                                      v
                          +-------------------------+
                          |   3. DECIDE             |
                          |   LLM-grounded synthesis|   <- gov-cloud LLM
                          |   with citation         |      (GPT-4o / Gemini /
                          |   enforcement           |       Grok / corpus-only)
                          +-----+-------+-----------+
                                |       |
                       low-confidence   |  high-confidence
                                |       |
                                v       v
        +----------------+   +-------------------------+
        |  HUMAN ON THE  |   |   4. ACT                |
        |  LOOP HOLD     |   |   emit to consumer:     |
        |  (operator     |   |   Foundry action,       |
        |  approve/deny) |   |   AIP tool call,        |
        +-------+--------+   |   NGIC2 widget,         |
                |            |   MSS workflow,         |
                |            |   CoT to TAK / JBC-P    |
                |            +-----------+-------------+
                |                        |
                +------------------------+
                                      |
                                      v
                          +-------------------------+
                          |   5. RECORD             |
                          |   AI audit engine logs  |
                          |   provider, model,      |
                          |   prompt hash, citation |
                          |   chain (DoD AI gov.)   |
                          +-----------+-------------+
                                      |
                                      v
                          +-------------------------+
                          |   6. LEARN              |
                          |   RAG self-improvement  |
                          |   loop indexes output   |
                          |   back into corpus,     |
                          |   backtest scores prior |
                          |   forecasts vs ground   |
                          |   truth weekly          |
                          +-----------+-------------+
                                      |
                                      +--> back to OBSERVE

Build-vs-buy — the math the program office needs

Capability MSS / AIP / Foundry need to buildRealistic build costStatus in ATLAS today
OSINT collector swarm across 8+ sources (Telegram, Mastodon, Bluesky, YouTube, RU/CN media, GDELT, procurement, cyber)~6 dev-months per source × 8 = ~48 dev-monthsSHIPPED (osint-collectors, every 15–60 min)
Source-provenance classifier per NATO STANAG 2511 (reliability A–F × credibility 1–6)~3 dev-months + doctrine reviewSHIPPED (intel-source-provenance, on collection)
LLM-grounded synthesis with hard citation enforcement (no [src] = no assertion)~6 dev-months including eval harnessSHIPPED (intel-synthesis, every 6h)
DIME / PMESII-PT briefing emitter with deep-link provenance~4 dev-monthsSHIPPED (intel-briefing)
Drone threat-evolution ledger (Geran-3, Yolka, Ptashka, Lancet, Shahed family)~3 dev-months ongoing curationSHIPPED (intel-tag-watcher + ledger)
Kill-chain bridge: synthesis → CoT → TAK / JBC-P~4 dev-monthsSHIPPED (intel-killchain-bridge)
Sensor fusion + WEZ geometric engine~9 dev-months including geodesy reviewSHIPPED (sensor-fusion, wez-engine)
Government-cloud LLM orchestrator with failover, audit, citation tagging~5 dev-monthsSHIPPED (ai-provider-orchestrator + ai-audit)
RAG self-improvement loop + weekly backtest harness~4 dev-monthsSHIPPED (intel-rag-loop, intel-backtest)
Daily Commander Update Brief auto-assembler~2 dev-monthsSHIPPED (intel-cub, daily 0500Z)
SAM.gov opportunity watcher with auto-classification~2 dev-monthsSHIPPED (sam-watcher)
Voice tactical translator (mic → CoT) for hands-busy ops~3 dev-monthsSHIPPED (voice-tactical)
Swarm fail-safe + Reconnection Guard (DDIL safe-to-reconnect posture check)~3 dev-monthsSHIPPED (swarm-failsafe + reconnection-guard)
Total honest comparable build cost for a Government program office to recreate the agent set in MSS / AIP / Foundry: roughly 50–60 dev-months across 8–12 specialists, 18–24 calendar months, plus doctrine review, plus the OSINT-source maintenance burden (which never ends). All of it is shipped today. Reviewer can call any endpoint in /agents with curl in 30 seconds and verify.

How ATLAS plugs into MSS / AIP / Foundry — three concrete patterns

ConsumerHow it consumes ATLASWhat the operator sees
Palantir AIPAIP registers each agent's OpenAPI spec from /api/atlas/foundry-actions.json as a native tool. AIP's logic builder calls them like any other tool.AIP agents in the operator's existing AIP UI gain access to live DIME/PMESII synthesis, drone-evolution ledger, and OSINT-grounded answers. No new GUI.
Palantir FoundryFoundry actions wrap each ATLAS endpoint. Foundry ontologies inherit the agent's outputs as columns or linked objects.Analyst inside Foundry sees the latest ATLAS synthesis as part of their existing object workflow. They never know it came from outside.
MSS workflowsMSS workflow steps make HTTP calls to ATLAS endpoints (same as AIP). Outputs feed the next MSS step.MSS-driven mission threads inherit ATLAS agent output as workflow context. The agentic AI workflow the warfighter SME described, with ATLAS providing the agents.

Representative agents (live, callable now)

AgentCycleConsumesEmits toEndpoints
Intel Synthesis Cycle every 6 hours upstream feeds Threat-evolution ledger, briefing emitter, kill-chain bridge /api/intel/synthesis/latest
/api/intel/synthesis/history
Kill-Chain Bridge on intel update Intel synthesis TAK Bridge, CoT Translator, TAK Subscriber /api/intel/killchain-bridge
/api/cot/dispatch
OSINT Collector Swarm every 15 to 60 minutes per source upstream feeds operator console + downstream consumers /api/osint/health
/api/osint/delta/stream
/api/osint/corroboration
Sensor Fusion Orchestrator real time, sub-second upstream feeds operator console + downstream consumers /api/sensors/fusion
/api/sensors/triangulation
Weapons Engagement Zone (WEZ) Engine on track update upstream feeds operator console + downstream consumers /api/wez/zones
/api/wez/recommend
TAK Bridge real time upstream feeds operator console + downstream consumers /api/tak/status
/api/cot/dispatch
SAM.gov Opportunity Watcher daily upstream feeds operator console + downstream consumers /api/procurement/opportunities
/api/sbir/proposals
AI Audit Engine on every AI call upstream feeds operator console + downstream consumers /api/ai-audit/recent
/api/ai-audit/spend

Full catalog of 20 agents at /agents. Machine-readable Foundry/AIP tool spec at /api/atlas/foundry-actions.json.

Human-on-the-loop, not human-in-the-loop

Per the warfighter SME's third comment in the same thread — "direction of the industry is end-to-end agentic AI" — the goal is to remove the human from every routine step while keeping them on the loop for the moments that matter. ATLAS enforces this as follows:

JSON specs: GET /api/atlas/agents · GET /api/atlas/foundry-actions.json · ATLAS Posture v1.0.0-2026.05.20 · Direct answer to the warfighter SME CAPES Slide 3 comment, 28 Apr 2026.
PRESENCE ✓
⚠ SANDBOX / TRAINING MODE — Live read-only data. Write commands are inhibited (train as you fight, missile button safed).