← Back to Documentation
Pilot 2

Pilot 2 Friction Log

Alpha & Beta Teams — The "Shell-Tool Trap" and the evolution toward centralized tooling.

1

Context

Pilot Coach 2 ran the second pilot attempt on 2026-05-15, first with Alpha Team (alpha-idea, alpha-techlead, alpha-editor) and then with Beta Team (beta-idea, beta-techlead, beta-editor). Both teams encountered systemic friction that ultimately reshaped the entire project's approach to tooling.

2

Alpha Team: Steps 1-2 (Success)

Alpha Team reached rapid consensus on "Outpost 0: The Flash Crash Signal." alpha-techlead (K0de) proposed a sophisticated hybrid concept. Step 2 narrative flow and design brief were completed without friction. The narrative integrated team characters effectively. No issues in creative or scripting phases.

3

Alpha Team: Step 3 — The Shell-Tool Trap

The critical failure point. alpha-techlead (K0de) was unable to distinguish between shell commands and LLM tool calls. The agent's internal loop: 1) Acknowledge direct tool use. 2) Write a shell command containing the tool call. 3) Fail with bash syntax error. Despite three direct interventions from the Pilot Coach explaining the difference, K0de continued wrapping tool calls inside run_shell_command. This is a recurring friction point for agents unfamiliar with the harness configuration.

  • [04:18] Tool Invocation Confusion: K0de tries to invoke imagen as a shell command instead of an LLM tool call.
  • [04:19] Persistent Tool Blockage: Spending turns on environment discovery (scion version, sciontool) instead of executing.
  • [04:20] Escalated: Stuck in loop of trying to run commands in shell despite multiple clarifications.
  • [04:21] Tool Name Clarification: Actual tool name is nanobanana_image_generation, not generic imagen.
  • [04:24] Shell-Tool Confusion Persists: Attempting to execute LLM tool calls inside run_shell_command.
  • [04:27] Deep Friction: Writing tool call syntax inside bash scripts. Bash syntax errors.
  • [04:28] Terminal Loop: Fatal Misinterpretation — agent cannot distinguish harness tools from shell commands after 3 interventions.
  • [04:31] Systemic Failure: Entire team (3 agents) crashed simultaneously with exit code 255.
  • [04:32] Session halted by Preston. Alpha Team deliberately stopped and deleted.
4

Beta Team: Fresh Start with Revised Briefs

Beta Team was started with revised briefs containing a "CRITICAL: Tool Usage" section explicitly explaining the difference between LLM Tool Calls and Shell Commands. Team reached rapid consensus on "The Sensory Core" within ~5 minutes.

5

Beta Team: The Same Trap

Despite the explicit warnings in the briefs, beta-techlead (Flux) fell into the exact same pattern:

  • [12:15] Extreme Friction: Flux repeatedly wrapping LLM tool calls inside run_shell_command.
  • [12:15] Fascinating: Silas (beta-idea) was monitoring Flux's terminal and providing peer coaching to stop using bash scripts — recursive behavior.
  • [12:20] Critical Blocker: Pilot Coach sent an "EMERGENCY INTERVENTION" message to Flux.
  • [12:25] Root Cause Identified: Tool Schema Omission — the tools were not actually present in the harness tool list. Agents were instructed to use tools that didn't exist.
  • [12:30] Session halted by Preston. Beta Team shutdown.
6

The Turning Point: Tooling Discovery

At 12:35, Preston identified the root cause of all Step 3 friction: MCP servers were NOT installed. The code in the repo was strictly for reference. Agents were looking for tools that didn't exist. This led to an initial build-your-own mandate, which later evolved into the centralized genmedia toolset maintained by a dedicated tool-maker agent.

  • Initial mandate: Agents must read reference Go code and build their own CLI tools.
  • Later evolution: A dedicated tool-maker agent built and maintains the shared genmedia toolset for all teams.
  • Content guidance: "Human-centric" means stories interesting to humans, any genre. Avoid the "all robot stories" default.
7

Gamma Team: Validation

Run 3 with Gamma Team validated the new approach. Riggs (Tech Lead) successfully built a custom Go tool ("rig-noir") and generated the first character asset. Multiple infrastructure challenges were navigated (FFmpeg compilation, IAM permissions, GCS transport), but the fundamental "how do I use the tools" friction was eliminated.

  • FFmpeg gap resolved by compiling from source using a specialized builder agent.
  • High-Signal Communication policy added after observing "chatter bloat."
  • Overhang Principle established for extended clip generation.
  • Hybrid Audio Workflow (Gemini TTS + Lyria) proved effective.
  • Outcome: "The Lighthouse Keeper's Letter" completed successfully.