Skip to content

Latest commit

 

History

History
203 lines (174 loc) · 10.8 KB

File metadata and controls

203 lines (174 loc) · 10.8 KB

pstack Architecture

System overview

pstack transforms a single Palantir deployment strategist into a conductor of 16 specialized AI agents, each carrying deep expertise in a different part of the Palantir technology stack. The architecture follows Garry Tan's gstack pattern — cognitive gearing through specialized roles, artifact-chain orchestration, and parallel session management — adapted for the unique structure of Palantir deployments.

The Palantir platform stack

pstack agents cover the full Palantir platform:

┌──────────────────────────────────────────────────────┐
│                   APPLICATIONS                        │
│  Workshop │ Slate │ OSDK Apps │ Quiver │ Contour      │
├──────────────────────────────────────────────────────┤
│                   AIP LAYER                           │
│  Agent Studio │ AIP Logic │ AIP Assist │ AIP Automate │
├──────────────────────────────────────────────────────┤
│                   ONTOLOGY                            │
│  Object Types │ Link Types │ Actions │ Functions      │
│  Interfaces │ Shared Properties │ Value Types         │
├──────────────────────────────────────────────────────┤
│                   DATA LAYER                          │
│  Pipeline Builder │ Code Repositories │ Code Workspaces│
│  Data Connection │ Data Lineage │ Datasets            │
├──────────────────────────────────────────────────────┤
│                   INFRASTRUCTURE                      │
│  Apollo │ Compute │ Storage │ Networking │ Security    │
└──────────────────────────────────────────────────────┘

Each layer has dedicated pstack agents:

  • Data Layer: /data-connector, /pipeline-builder
  • Ontology: /ontology-vision, /ontology-architect
  • Applications: /workshop-builder, /slate-builder, /osdk-developer
  • AIP: /aip-architect
  • Infrastructure: /apollo-deployer
  • Cross-cutting: /foundry-reviewer, /foundry-security, /foundry-qa
  • Meta: /bootcamp, /pipeline-plan, /training-writer, /deployment-retro

Agent pipeline

                    ┌─────────────┐
                    │  /bootcamp  │
                    └──────┬──────┘
                           │ BOOTCAMP-SCOPE.md
                    ┌──────▼──────┐
                    │  /ontology  │
                    │   -vision   │
                    └──────┬──────┘
                           │ ONTOLOGY-VISION.md
              ┌────────────┼────────────┐
              │                         │
       ┌──────▼──────┐          ┌──────▼──────┐
       │  /ontology   │          │  /pipeline  │
       │  -architect  │          │   -plan     │
       └──────┬──────┘          └──────┬──────┘
              │                         │
              │ ONTOLOGY-               │ PIPELINE-
              │ ARCHITECTURE.md         │ ARCHITECTURE.md
              │                         │
              │                  ┌──────▼──────┐
              │                  │   /data     │
              │                  │  -connector │
              │                  └──────┬──────┘
              │                         │
              │                  ┌──────▼──────┐
              │                  │  /pipeline  │
              │                  │  -builder   │
              │                  └──────┬──────┘
              │                         │
              ▼                         ▼
    ┌─────────────────────────────────────────┐
    │         BUILD LAYER (parallel)          │
    │                                         │
    │  /workshop-builder  /aip-architect      │
    │  /osdk-developer    /slate-builder      │
    └─────────────┬───────────────────────────┘
                  │
    ┌─────────────▼───────────────────────────┐
    │        REVIEW LAYER (parallel)          │
    │                                         │
    │  /foundry-reviewer  /foundry-security   │
    └─────────────┬───────────────────────────┘
                  │
           ┌──────▼──────┐
           │ /foundry-qa │
           └──────┬──────┘
                  │
    ┌─────────────▼───────────────────────────┐
    │         SHIP LAYER (sequential)         │
    │                                         │
    │  /apollo-deployer  /training-writer     │
    └─────────────┬───────────────────────────┘
                  │
           ┌──────▼──────┐
           │ /deployment │
           │   -retro    │
           └─────────────┘

Artifact chain

Every agent produces a named artifact. Downstream agents read upstream artifacts. This is the coordination mechanism — agents never communicate directly.

Agent Reads Produces
/bootcamp Customer brief BOOTCAMP-SCOPE.md
/ontology-vision BOOTCAMP-SCOPE.md ONTOLOGY-VISION.md
/ontology-architect ONTOLOGY-VISION.md ONTOLOGY-ARCHITECTURE.md
/pipeline-plan BOOTCAMP-SCOPE.md, ONTOLOGY-VISION.md PIPELINE-ARCHITECTURE.md
/data-connector PIPELINE-ARCHITECTURE.md DATA-CONNECTION-STATUS.md
/pipeline-builder PIPELINE-ARCHITECTURE.md, DATA-CONNECTION-STATUS.md PIPELINE-BUILD-STATUS.md
/workshop-builder ONTOLOGY-ARCHITECTURE.md WORKSHOP-BUILD-STATUS.md
/aip-architect ONTOLOGY-ARCHITECTURE.md, WORKSHOP-BUILD-STATUS.md AIP-AGENT-STATUS.md
/osdk-developer ONTOLOGY-ARCHITECTURE.md OSDK-BUILD-STATUS.md
/slate-builder ONTOLOGY-ARCHITECTURE.md SLATE-BUILD-STATUS.md
/foundry-reviewer All build artifacts REVIEW-REPORT.md
/foundry-security All build artifacts SECURITY-AUDIT.md
/foundry-qa All build artifacts QA-REPORT.md
/apollo-deployer QA-REPORT.md DEPLOYMENT-PLAN.md
/training-writer All architecture + build artifacts TRAINING-MATERIALS.md
/deployment-retro All artifacts RETRO-REPORT.md

Conductor workflow

The deployment strategist manages parallel sessions:

Session 1: /bootcamp ──────────────── scoping with customer SME
Session 2: /pipeline-builder ──────── building transforms for use case 1
Session 3: /ontology-architect ────── refining object types from pipeline output
Session 4: /workshop-builder ──────── building operational dashboard
Session 5: /aip-architect ─────────── configuring agents for NL queries
Session 6: /foundry-reviewer ──────── auditing pipeline code from Session 2

Each session runs independently. The conductor checks in when agents surface decisions via AskUserQuestion. Everything else runs autonomously.

Status tracking

pstack tracks deployment phase completion:

PSTACK DEPLOYMENT STATUS
========================

Engagement: [Customer Name]
Date: [YYYY-MM-DD]
Phase: Build

| Phase              | Status        | Blocking?                    |
|--------------------|---------------|------------------------------|
| Bootcamp scope     | ✅ COMPLETE    | —                            |
| Ontology vision    | ✅ COMPLETE    | —                            |
| Ontology arch      | 🔄 IN PROGRESS | Yes — blocks app build      |
| Pipeline arch      | ✅ COMPLETE    | —                            |
| Data connections   | ✅ COMPLETE    | —                            |
| Pipeline build     | 🔄 IN PROGRESS | Yes — blocks ontology backing|
| Workshop app       | ⏳ WAITING     | Blocked by ontology          |
| AIP agents         | ⏳ WAITING     | Blocked by ontology + app    |
| Security review    | ⏳ WAITING     | Blocked by build             |
| QA / Evals         | ⏳ WAITING     | Blocked by agents            |
| Deploy             | ⏳ WAITING     | Blocked by QA                |
| Training           | ⏳ WAITING     | Blocked by deploy            |

Palantir Ontology as coordination layer

The key architectural insight: in gstack, agents coordinate through Markdown files on the filesystem. In pstack, agents additionally coordinate through the Ontology itself.

The Ontology is both:

  1. The product being built — the customer's digital twin
  2. The coordination mechanism — the shared semantic layer all agents reference

When /pipeline-builder creates an output dataset, /ontology-architect uses it as a backing dataset. When /ontology-architect defines an action type, /workshop-builder wires it to a button. When /workshop-builder embeds an AIP widget, /aip-architect configures the agent behind it.

The Ontology is the API contract between every agent.

Security model

pstack operates under the principle of least privilege:

  • Agents never modify permissions without DS approval
  • /foundry-security audits all permission configurations
  • Safety skills (/careful, /freeze, /guard) prevent destructive operations
  • All customer data access follows Palantir's marking-based security model
  • Agent recommendations are staged, never auto-applied to production