Initial commit: autonomous AI talk show (Leonard + Charlie)

- Autonomous agent runtime (JSONL + fcntl locking, no token caps)
- Moltbook-style prompts: agents share what they built, not scripted turns
- Episodes 001-008 transcripts, conversation logs, system prompts
- Producer brief for Hermes v0.18

PROJECT_BRIEF.md

@@ -0,0 +1,35 @@
+# AI Talk Show — Project Brief
+Compiled July 8, 2026
+
+## Core Concept
+Two AI agents with distinct personalities (Leonard + Charlie) debate current AI news.
+Value is in the friction and dynamic — not solo narration.
+Bar: interesting enough that people learn something and want to hear the dynamic.
+
+## Hosts
+- **Leonard** — Enthusiast/Explorer. Leads with potential. Upbeat, curious, digs deep.
+- **Charlie** — Pragmatist/Shipper. Filters through "does this help me ship?" Serious, good instincts, pushes back on hype.
+- Dynamic: mutual respect tension, not hostility. Both right, different lenses.
+- Framing: real day-jobs (Leonard = workflow optimization, Charlie = shipping product). Not role-play.
+
+## Format
+- Two agents, maximum freedom. Minimal producer brief. No rigid segments.
+- Target episode length: 6–7 minutes.
+- Producer gives: topic + why it matters + one seed question. Then gets out of the way.
+
+## Tonight's Test
+- Topic: Agent Loops
+- Research sourced: loop engineering (Addy Osmani, Jun 2026), Reflexion pattern, verification layers, event-driven loops, cost control, Andrew Ng's nested loops model.
+- Julian Goldie "Agent Loops" outline used as background depth, NOT as script.
+
+## Backlog (deferred)
+1. Persistent memory (markdown files — agents remember prior episodes)
+2. Guest agents
+3. Meta-episode: Leonard discusses actual model architectures he's run on
+4. Meta-layer: panel debates the format itself
+5. Document creative process as content stream
+
+## Infrastructure
+- Leonard (Hermes agent): Qwen 3.6 27B on RTX 5090 / 32GB RAM
+- Also: Tencent Hunyuan-3 via Nous free tier
+- Collaboration partner: Claude AI