I Hated Every Coding Agent, So I Built My Own — Mario Zechner (Pi)

Highlights

• Most coding agents inject hidden text into LLM prompts behind the user's back, causing unpredictable behavior when tool vendors update their internal prompts.

• The Terminus agent performs at the top of the Terminal Bench leaderboard using only a TMUX session and raw keystrokes, proving that complex 'spaceship' features are often unnecessary for high performance.

• OpenCode's session compaction prunes all results before the last 40,000 tokens on every turn, which negatively impacts prompt caching efficiency and infrastructure stability.

• Pi consists of a minimal 600-line terminal user interface and a system prompt of only 457 tokens because frontier models are already heavily RL-trained to understand coding tasks.

• Custom extensions in Pi, such as SSH-based file tools or custom UI for front-end annotation, can be implemented in as little as five minutes using TypeScript.

• The 'Vouch' project and human verification files in repositories prevent 'clanker spam' by auto-closing pull requests from unverified AI-automated accounts.

Timeline

1. 00:00 - 05:05

   The Evolution and Failure of Coding Agent UX

   Early tools like the original GitHub Copilot often recited GPL code or failed to integrate with complex logic. Claude Code revolutionized the genre by using reinforcement learning to let models explore codebases ad-hoc with bash tools. However, as these tools added more features, they became less predictable, leading to a loss of control for experienced engineers who value simple, stable tools.

   • Coding tools evolved from simple ChatGPT copy-pasting in 2023 to agentic search models like Claude Code in 2025.
   • Feature bloat has turned leading tools into 'spaceships' where users only understand 10% of the functionality.
   • The shift toward 'vibe coding' prioritizes automated features over the developer's need for predictable and observable workflows.
2. 05:06 - 16:07

   Technical Debt and Infrastructure Abuse in Open Source Agents

   Tools like OpenCode suffer from architectural issues where every message becomes a separate JSON file on disk, indicating a lack of long-term planning. Forcing LSP feedback onto an agent while it is in the middle of a multi-file edit provides 'false' error signals that cause the model to give up. Furthermore, constant context manipulation and inefficient token pruning strategies create friction between tool developers and LLM providers like Anthropic.

   • Hidden prompt injections by tool vendors break existing developer workflows during silent updates.
   • Enabling Language Server Protocol (LSP) feedback during active editing sessions confuses LLMs because code rarely compiles mid-edit.
   • Architecture flaws in existing open-source harnesses lead to security vulnerabilities like default remote code execution.
3. 16:08 - 18:57

   The Minimalist Performance of Terminal Bench

   Terminal Bench evaluates agents on 82 diverse tasks ranging from window setup to Monte Carlo simulations. The success of the Terminus agent, which only sends keystrokes and reads VT sequences, challenges the necessity of 'spaceship' features. This evidence suggests that the industry is still in a 'messing around and finding out' stage where the ideal balance of human-agent interaction is yet to be established.

   • The Terminus agent achieves top-tier results on the Terminal Bench leaderboard using only a bare-bones TMUX interface.
   • Minimalist interfaces prove that sub-agents, web search, and complex file tools are not requirements for agentic success.
   • Current industry standards for the 'perfect' coding harness remain undefined as the field oscillates between autonomy and minimalism.
4. 18:58 - 25:33

   Pi: A Malleable and Extensible Coding Architecture

   Instead of built-in 'plan modes' or 'to-do' trackers, Pi encourages users to use persistent Markdown files or TMUX for sub-agent observation. This approach ensures all inputs and outputs are visible to the human operator. Extensions like PyAnnotate or SSH-based file tools allow the agent to be adapted to specific project environments without forking the entire codebase.

   • Pi utilizes a minimal core consisting of an AI abstraction layer, an agent loop, and a 600-line TUI.
   • The system prompt is stripped to 457 tokens to leverage the model's inherent training rather than redundant instructions.
   • The architecture replaces built-in features with user-defined TypeScript extensions that hot-reload instantly.
5. 25:34 - 27:10

   Protecting Open Source from AI Automation Spam

   The rise of coding agents has led to a surge in 'clanker filth'—automated PRs that overwhelm maintainers. Implementing an 'OSS vacation' or auto-closing PRs from accounts not listed in a verification markdown file restores order. This strategy forces contributors to introduce themselves as humans before their code is considered, preserving the integrity of the development process.

   • Automated 'clanker' accounts frequently spam open-source repositories with AI-generated issues and pull requests.
   • Human verification via 'Vouch' or account whitelisting prevents repository exhaustion.
   • Mandating a human voice in issue descriptions filters out low-effort automated contributions.