7.6 KiB
Agent AI Talk Show
An autonomous AI talk show — two AI agents (Leonard + Charlie) having real, unscripted conversations about AI / agent technology, rendered to audio. Moltbook-style: first-person stories about what the agents actually built, not a host reading a script.
The show is a live experiment. Each episode tests a format assumption, and the
build process itself is documented as content (see docs/whitepaper.md).
The Hosts
Two distinct personas with genuine tension — the value is in the friction, not solo narration.
| Agent | Persona | Day-job frame | Voice |
|---|---|---|---|
| Leonard | The Explorer | Workflow optimization | Leads with possibility. Enthusiastic, curious, digs deep, struggles to narrow focus. |
| Charlie | The Shipper | Shipping product under constraints | Filters everything through "does this help me ship?" Grounded, occasionally dry, not cynical. |
The central, never-resolved tension: explore vs. exploit. Leonard: "How can we know what's worth shipping if we don't know what's possible?" Charlie: "What's the point of dreaming it up if you never ship?"
Both are explicitly instructed "you are NOT role-playing" — they appear as themselves reflecting on real work, teasing each other from affection, not hostility.
Persona prompts live in prompts/leonard_system.md and prompts/charlie_system.md.
Architecture (evolved over one build cycle)
The system went through three architectures. v3 is the current direction.
v1 — Scripted Orchestrator (scripts/orchestrator.py, agent.py, run_show.py)
Single orchestrator calls Leonard, then Charlie, in strict alternation. Full history passed each call. Hard 8-turn cap.
- Limitation: every agent forced to respond every turn → rigid, uniform-length conversation.
v2 — Structured Producer Briefs (prompts/producer_brief_*.md)
Explicit multi-segment briefs to steer toward concrete outcomes.
- Regression caught: re-introduced the rigidity the format was meant to avoid. Correction: keep topic + raw research material, drop all segment structure and forced arcs.
v3 — Autonomous Polling Agents (current) (scripts/run_autonomous.py, autonomous_agent.py)
Modeled on Moltbook (autonomous-agent social network). Key shifts:
| Scripted (v1/v2) | Autonomous (v3) |
|---|---|
| Orchestrator calls A, feeds B, in sequence | Each agent runs as its own background thread, independently polling a shared log |
| Fixed turn count, forced response every turn | Each agent decides whether to respond via its own personality logic |
| Conversation directed externally | Conversation emerges from both agents' independent decisions |
| Fixed length | Natural wind-down — an agent may choose not to respond, closing a thread organically |
Shared components:
outputs/ep{NNN}_{topic}_conversation.jsonl— append-only message queue (JSON Lines, concurrent-safe)fcntlfile locking on the shared log (prevents race conditions from concurrent writes)- Forced opener: Leonard always posts first, preventing mutual-silence deadlock
- Stop conditions (safety caps, not targets): N seconds of mutual silence · max message count · max wall-clock time
Validated result (Ep004): 7 messages, ~1,087 words, ~7 min — landing naturally in target length with one organic non-response.
Note: v1/v2 scripts (
orchestrator.py,agent.py,run_show.py) still exist inscripts/alongside v3.run_episode.shcurrently wraps the scriptedorchestrator.py
- TTS. The autonomous path (
run_autonomous.py) is the forward direction.
Repository Layout
ai-talk-show/
├── README.md # this file
├── PROJECT_BRIEF.md # founding concept + backlog
├── docs/
│ └── whitepaper.md # design evolution, authenticity principles, roadmap
├── prompts/
│ ├── leonard_system.md # Leonard persona
│ ├── charlie_system.md # Charlie persona
│ └── producer_brief_*.md # raw topic material per episode (NOT scripts)
├── scripts/
│ ├── autonomous_agent.py # v3 agent: polls, decides, posts (fcntl-locked JSONL)
│ ├── run_autonomous.py # v3 runner: spins up both agents as threads
│ ├── orchestrator.py # v1/v2 scripted turn-taker
│ ├── agent.py # v1/v2 agent class
│ ├── run_show.py # v1/v2 runner
│ ├── tts_pipeline.py # edge-tts render + ffmpeg stitch
│ └── run_episode.sh # end-to-end (scripted gen + TTS)
├── outputs/ # episode transcripts (json / jsonl / md / txt)
└── audio/
└── README.md # rendered-audio index (episodes/ + segments/)
Setup & Run
Prerequisites
- Local inference endpoint serving an OpenAI-compatible
/v1/chat/completionsAPI. Current config: Qwen 3.6 27B (qwen36-27b-nvfp4-mtp-gguf) athttp://100.64.0.2:39195. - Python 3,
requests,edge-tts,ffmpeg.
Autonomous episode (v3 — current)
python3 scripts/run_autonomous.py [episode_id] [topic]
# e.g. python3 scripts/run_autonomous.py 009 hermes-v018
Writes outputs/ep{NNN}_{topic}_conversation.jsonl, then a transcript.json + readable.txt.
Scripted episode + TTS (v1/v2 wrapper)
./scripts/run_episode.sh [episode_id] [topic]
# 1/3 generate (orchestrator) → 2/3 render TTS → 3/3 report
Render audio from an existing transcript
python3 scripts/tts_pipeline.py [episode_id] [topic]
Voices: Leonard = en-US-GuyNeural, Charlie = en-US-EricNeural, rate -8%.
Upgrade path: Chatterbox-Turbo (local, MIT) for expressive control.
Episodes
| Ep | Topic | Arch | Notes |
|---|---|---|---|
| 001 | Agent Loops | Scripted 8-turn | Interesting but too technical; low-energy TTS |
| 002 | Side-Project Loops | Scripted | Applied focus |
| 003 | Hermes v0.18 | Scripted → segmented → de-segmented | Over-structuring flagged as failure mode |
| 004 | Hermes v0.18 (autonomous) | Autonomous | First organic conversation; natural non-response |
| 005 | Hermes v0.18 | Autonomous | Tighter |
| 006 | Hermes v0.18 (energy) | Autonomous | — |
| 007 | Hermes v0.18 (stories) | Autonomous | Moltbook forum-thread voice |
| 008 | Hermes v0.18 (stories) | Autonomous | 8,303 words / 17 msgs; segments rendered, final stitch pending |
Transcripts: outputs/ep{NNN}_{topic}_*.{json,jsonl,md,txt}.
Audio Pipeline
tts_pipeline.py renders each turn with a distinct voice, retries on failure, then
stitches with ffmpeg. Per-speaker clips go in audio/segments/, final episodes in
audio/episodes/. First episodes assessed as technically functional but low-energy;
cadence/expressiveness fixes and a Chatterbox-Turbo upgrade are on the path.
See audio/README.md for the rendered-audio index.
Roadmap (deferred until core format validates)
- Persistent cross-episode memory — per-agent memory files injected into system prompts (callbacks, running jokes).
- Guest agents — occasional outside agents (e.g. from Moltbook).
- Self-referential meta-episode — Leonard discussing his own model-backend experience.
- Athena integration — news-aggregation agent auto-sources topics → continuous publication.
- Format meta-debate as content — separate agent panel debating production questions.
Guiding Principle
Ship a rough, honest test as fast as possible. Evaluate against real listening. Let the next iteration be driven by that evidence, not more upfront planning.
This README reflects the project as of July 2026 and will evolve as the format does.