Agentic Infrastructure — Blueprint

Autonomous Agents That
Operate Your Product

AI search is table stakes. The difference is the infrastructure that runs it — intent routers that classify fan queries, delegation layers that dispatch to specialized agents, autonomous loops that self-improve embeddings from real usage data, and MCP bridges that connect to Storyteller's SDK without fragile API integrations.

The Architecture

Multi-Agent Orchestration Layer

Four agent types work together. Each has a specific job. The orchestration layer routes work, delegates tasks, and monitors results — without human initiation.

Intent Router

Hermes — 136+ Specialized Skills

Classifies every fan query by type: stat lookup, narrative search, comparison, timeline, highlight reel. Routes to the correct agent. Handles ambiguous queries by spawning parallel searches and merging results.

↓ delegates to ↓
Delegation

Paperclip — Parallel Agent Dispatch

Spawns specialized sub-agents for each query dimension. "LeBron vs Boston playoffs" triggers: a stat agent, a narrative agent, a video highlight agent. All three run simultaneously. Results merge before the fan sees anything.

↓ self-improves via ↓
Autonomous Loop

Agent Zero — Continuous Embedding Refinement

Monitors which search results fans click. Identifies gaps where no good story exists. Proposes new embedding strategies. Updates the vector index without human intervention. FAISS vector memory retains learnings across sessions.

↓ connects via ↓
MCP Bridge

MCP Servers — SDK, Stats, Content APIs

Connects agents to Storyteller's SDK, NBA stats APIs, and content libraries. Agents call these tools directly — no middleware, no manual API integration per data source. New data sources are added as MCP servers, not as code changes.

The Difference

What This Infrastructure Does That a Search Bar Cannot

1

Autonomous Content Gap Detection

Agent Zero monitors failed searches — queries where no story ranked above threshold. It logs the gap, checks if the content exists elsewhere in the library with different metadata, and flags what's truly missing. The editorial team gets a daily gap report they never asked for.

2

Self-Improving Embeddings

The system learns which embedding strategies produce the highest click-through. A "Giannis dunk" query that fans ignore gets re-weighted. A "LeBron clutch" query format that drives engagement gets propagated across all player pages. No ML team required.

3

Cross-League Knowledge Transfer

What the system learns from NBA fan behavior — which query types work, which embeddings perform — transfers to NFL, MLB, and WNBA deployments automatically. The agentic infrastructure is league-agnostic. Only the content index changes per client.

4

Immutable Audit Trail

Every search result, every agent decision, every embedding update logged with SHA-256 hash chaining. Content partners can verify their stories are being surfaced correctly. No black-box algorithm. Every result traceable to its source story and the agent that selected it.

Production Capabilities

What Ships With This Infrastructure

I
Intent Classification
136+ skill router classifies every query by type before processing
II
Parallel Agent Execution
Complex queries spawn 3-5 specialized agents running simultaneously
III
Autonomous Monitoring
Watchdog detects failures, auto-recovers with exponential backoff
IV
Content Gap Reports
Daily automated brief on what fans searched for and could not find
V
Embedding Self-Optimization
Click-through data feeds back into embedding weights automatically
VI
Cross-Client Transfer
Learnings from one league deploy to the next without re-engineering
VII
SHA-256 Audit Trail
Every agent action logged with hash-chained verification
VIII
MCP Server Integration
New data sources connect as servers, not as code changes
How I Build

Four Weeks. One Sports Client. Agentic From Day One.

Week 1 — Source Audit & Agent Design

Every data source inventoried with authority classification. Storyteller SDK mapped. NBA stats APIs catalogued. Conflicts documented. Missing context flagged. Agent routing logic designed against verified source material, not assumptions.

Week 2 — Orchestration Layer

Intent router deployed. Parallel agent dispatch configured. MCP servers wired to Storyteller SDK and NBA APIs. Vector database seeded with story embeddings. Baseline metrics established for every query type.

Week 3 — Autonomous Loops

Agent Zero monitoring deployed. Content gap detection active. Embedding self-optimization running in shadow mode — logging recommendations, not applying them yet. Watchdog recovery configured with exponential backoff.

Week 4 — Ship, Measure, Self-Improve

Agentic search goes live on the sports client app. Embedding self-optimization graduates from shadow to active. Content gap reports begin delivering to editorial. A/B test measures: search engagement, story rediscovery, content gap closure rate. The system improves itself from here.