Coding Agent Comparison¶
Summary¶
A comparison of leading coding agents across architecture, features, and philosophy — from maximal platforms to minimal harnesses.
Agent Comparison Matrix¶
| Feature | Claude Code | Pi | Open Code | AMP | Factory Droid |
|---|---|---|---|---|---|
| Tools | 15+ | 4 (Read, Bash, Edit, Write) | Many | Many | Many |
| System prompt | ~14,000 tokens | <1,000 tokens | Large | Large | Large |
| Model choice | Anthropic only | Any provider | Anthropic only | Paid | Paid |
| Permissions | Modes (always-ask, yolo) | YOLO by default | Always-ask | Configurable | Configurable |
| MCP | Supported | Not built-in (extension) | Supported | Supported | Supported |
| Extensibility | Hooks (spawn processes) | TypeScript extensions | Limited | Limited | Limited |
| Sub-agents | Supported | Not built-in | Supported | Supported | Supported |
| Session branching | No | Yes (tree + fork) | No | No | No |
| Context management | Platform-controlled | User-controlled | Platform-controlled | Platform-controlled | Platform-controlled |
| Terminal Bench score | High | 6th place (pre-compaction) | High | High | High |
Key Tradeoffs¶
Maximal vs. Minimal¶
Maximal agents (Claude Code, Open Code) come with everything built in — convenient but opaque, with platform-controlled context that changes behind the scenes.
Minimal agents (Pi, Terminal benchmark harness) strip away everything unnecessary — more context window for actual work, full user control, but require more setup.
When to Use Each¶
| Scenario | Best Fit |
|---|---|
| Quick prototyping, Anthropic subscription | Claude Code |
| Full control, custom workflows, any model | Pi |
| Open-source preference, team use | Open Code |
| Maximum performance, budget available | AMP or Factory Droid |
| Local models, offline | Pi + Ollama |