Lisp Development with AI Agents: High Costs and Technical Challenges
AI Agent Performance with Lisp vs. Other Languages
A DevOps engineer using OpenRouter with Goose CLI for AI agentic development discovered significant differences in AI performance between programming languages. While working on an RSS reader format conversion tool in Lisp, he encountered multiple challenges that made development expensive and inefficient compared to Python or Go.
Technical Implementation Details
The engineer initially tried to have AI agents interact with the Lisp REPL using tmux commands: tmux capture-pane -t 0.0 -p | tail -n 1. This approach consumed excessive tokens, required sleep commands, and involved parsing tmux output. Claude showed some progress but still struggled, while cheaper models like DeepSeek and Qwen performed poorly despite working adequately for other tasks.
To improve the situation, he created tmux-repl-mcp, a Python tool that provides a more straightforward REPL interface. Instead of complex tmux interactions, agents could simply run execute_command in the REPL and receive output directly. He chose Python because his existing Goose configuration used uvx for tool installation and management.
Cost and Performance Comparison
The difference between Lisp and Python development with AI was dramatic. With Python, he could write all code and tests in a day or two using cheap models, requiring only semi-manual debugging. With Lisp, even after implementing tmux-repl-mcp, he spent $10 in 30 minutes with Claude, and the signal-to-noise ratio remained poor compared to Python.
Key observations from the experience:
- AI generates code on a path of least resistance, defaulting to common patterns like QuickLisp even when instructed to use alternatives like OCICL
- The high-latency request-response nature of AI APIs conflicts with REPL development workflows
- Languages with high internet volume (Go, Python) are orders of magnitude easier and cheaper for AI-assisted development
- AI has converted language popularity into real cost savings per million tokens
The engineer noted that regardless of language, his role remained similar: acting as an opinionated product owner managing the AI. However, the experience lacked the usual enjoyment of writing Lisp directly, leading him to consider rewriting the project in Go for better AI compatibility.
📖 Read the full source: HN AI Agents
👀 See Also
Building and Testing an MCP Server in Claude Desktop: Architecture and Lessons
A developer shares their experience building and testing an MCP server within Claude Desktop, detailing their architecture setup and practical lessons learned about tool schemas, debugging, and limitations.
MemAware benchmark shows RAG-based agent memory fails on implicit context retrieval
The MemAware benchmark tests whether AI agents can surface relevant past context when users don't explicitly ask for it, revealing that current memory systems score only 2.8% accuracy on hard implicit queries versus 0.8% with no memory.
Savecraft MCP Server Provides Claude with Accurate Magic: The Gathering Data
Savecraft is an open-source MCP server that parses MTG Arena Player.log locally, syncs game state, and gives Claude access to 12 expert reference modules built on real Magic: The Gathering data. The tool prevents Claude from hallucinating card names and rules by providing access to actual Arena data, draft recommendations from 17Lands, and the complete Scryfall database.
Rift: A Better Alternative to Git Worktrees with Instant Copy-on-Write Snapshots
Rift uses btrfs or APFS snapshots to create instant, space-efficient copies of Git repositories. Initialization, creation, and listing via CLI or JavaScript FFI.