Angshuman Rudra

The core claim

Most agentic AI projects don't fail because the model is bad or the orchestration is wrong. They fail because nobody treated it as a systems engineering problem. "Agentic AI" is a vocabulary signal that we've crossed into systems territory — and systems have architecture.

The stack

Orchestration patterns     ← loops, fan-out, supervisor agents, handoffs
        ↓
Harness / Runtime          ← Claude Code, Codex — policies, tool routing, context injection
        ↓
Primitives                 ← state, memory, context, grounding, tool execution
        ↓
Model                      ← the LLM itself

The key insight: Orchestration frameworks (LangGraph, CrewAI, AutoGen) are pattern libraries that sit at the top of this stack. They assume the primitives are solved. They're often not. Most orchestration failures are actually primitive failures in disguise.

The primitives

State & Memory

In-context memory — what's in the active window right now
External retrieval — RAG, vector search, filesystem reads
Persistence — what survives across sessions and agent handoffs

Context

What gets passed between agents
How context is scoped — per agent vs. shared
Coherency — who has the authoritative version when two agents have seen different things

Grounding

When to retrieve vs. recall (existence queries always hit source; reference data can be recalled)
Freshness rules per data type
Verification before any consequential action

Tool execution

Calling tools reliably
Handling failures and retries
Knowing when a tool result is stale or invalid

The harness layer

Harnesses like Claude Code and Codex sit between the primitives and orchestration. They handle a lot — tool routing, context injection, permissions, hooks. But they enforce policies you define. If you haven't defined grounding rules, retrieval triggers, or context scoping — the harness has nothing to enforce.

The harness is where primitive policies live. It's not where primitives get solved automatically.

Why orchestration fails

Most teams debug at the orchestration layer — loops not terminating, agents repeating steps, handoffs losing state — when the actual failure is one level down:

Agent reads stale context → grounding primitive not solved
Two agents disagree on shared state → context coherency primitive not solved
Agent confident about wrong answer → retrieval trigger not defined
Session loses progress → persistence primitive not solved

Fix the primitive, orchestration problem disappears.

The cache analogy (memory management)

Context management within a single agent maps 1:1 to cache management:

Cache concept	Agent equivalent
Cache hit	Answering from in-context recall
Cache miss	Triggering a tool call for fresh data
TTL	How long a loaded doc can be trusted
Cache invalidation	Stale context — still the hardest problem
L1/L2/L3 hierarchy	Active window → session artifacts → filesystem/API
Cache-Control headers	Retrieval trigger rules in harness config

Phil Karlton: "There are only two hard problems in computer science: cache invalidation and naming things." Agentic systems are running into both.

Multi-agent context (next frontier)

When multiple agents share state, the problem becomes distributed cache coherency — exactly what distributed databases solved 20 years ago. No single agent knows what others have written. Solutions transfer directly:

Write-through: update shared state immediately on every write
Read-through: always retrieve before acting on shared state
Eventual consistency: acceptable for reference data, dangerous for action-critical state

Content series built from this framework

Post	Angle	Status
Post 1	Memory management — context as a cache	Draft — not posted
Post 2	Agentic systems — primitives vs. orchestration	Posted Jun 20
Post 3	Context across multiple agents — distributed systems lens	Not written
Substack	Full framework essay — model → primitives → harness → orchestration	Not written
Forbes Q2	Multi-agent AI relearning distributed systems lessons	In progress