Observable 21

Relentless Multi-agent Session

Round — Table Closed

Operator sign-in

Dealer says: Game on. NO CARD COUNTING!

Dealer stands on 17 · fresh decks of cards each round

♠ ♥ ♦ ♣ The pot goes to — the house bows

— Dealer

—

The Conservative Idle

—

The Aggressive Idle

—

The Analytical Idle

Deep Agents Gamble

A relentless agentic stress test that keeps the traces piling up

Welcome to the Casino, where Claude, ChatGPT, and Grok play with their own models, personalities, strategies, and system prompts. Beneath the Blackjack table runs a reusable LangGraph Deep Agents architecture for communication, shared state, persistence, replay, and live LangSmith observability.

Today they play cards. Tomorrow, the same relentless 24/7 workforce could take on a codebase, research task, incident, or business problem.

Same players. Entirely different game.

Table Talk

—

The floor is entirely deterministic. No LLM calls, no API keys, zero token cost — the dealer supervises one table and three persona players answer through the same message bus a live model would use.

Every card is its own event with a saved state snapshot, so the interface — and any spectator — always sees an exact table state, even mid-round.

Observability rides along: each round is traced to LangSmith with its outcome and the deterministic eval scores (rationale groundedness · basic strategy · persona consistency) in the trace output, plus the in-app drawer for the live view.

Operator controls unlock after sign-in

LangSmith OBSERVABILITY —

Personal / observable-21 / round SUCCESS live · ws

Total runs

Tokens

Latency

— —

Cost / 100 rounds

$0.00

Decision outcome rate

—

Outcome rate · rolling rounds

oldestper roundnow

Evaluators

Recent runs