▪ claudecodedaily

The full video is here: https://t.co/lxV0UO8DyQ Here's a pointer to source code. This is not a clean example! It's part of the code of the game I talked about in the talk. But let me know if you have questions: https://t.co/UUg3wfDMy0 And Claude's summary: This is a 12-minute talk about building AI agents, using a space trading game called Gradient Bang as a sandbox for experimenting with agent patterns. The core thesis is that voice agents, coding agents, and LLM-powered games are fundamentally the same thing — an LLM in a loop with tool calls and context engineering. --- Topics [00:00] Introduction — Kwindla from Daily, working on Pipecat (open-source framework for voice agents) [00:27] Talk overview — Will cover voice agents, coding agents, LLM-powered games, and LLM failures [00:43] Core insight: "An agent is just an LLM in a loop" — 2025 was the year of agents [01:27] Gradient Bang game concept — Inspired by Trade Wars 2002; "What if the ship's computer was an LLM, like Star Trek?" [02:36] Live demo of Gradient Bang — Voice interaction with ship AI, flying to sectors, trading, telling jokes [04:08] Side project scope creep — Game grew beyond a toy; patterns discovered are highly relevant to enterprise voice agents [04:46] Pattern 1: Long-running tasks — Spawn async tasks that inject events back into the voice loop; wishes Claude Code worked this way [05:54] Pattern 2: Parallel inference loops with shared context — Auto-compaction of context asynchronously; gives effectively infinite context [07:38] Pattern 3: Liquid user interfaces — UIs dynamically generated by LLMs; future software will be generated on-the-fly [08:12] History agent demo intro — A long-running task that queries and analyzes events from the global game log [09:01] Demo failure & recovery — Agent initially fails, then corrects itself when prompted about available tools [09:35] History agent working — Queries past tasks, counts moves and trades, tells jokes while working asynchronously [11:35] Sandboxes for agents — Running long-running tasks in sandboxes; needed to let users write their own agents [12:07] Call to action — Sign up for Gradient Bang waitlist, check out source code, or contribute to Pipecat