Coding Agent Spec: Overview
Goal: Everything you need to build a working coding agent, in one place. Audience: Engineers who want to build, not just understand. Status: Reproduction-ready specification.
Design Principles
- Standalone - This folder contains everything. No external dependencies.
- Reproduction-ready - Complete specs, not summaries. You can build from this.
- Evidence-based - Derived from production agent analysis (Amp Code, Jan 2026).
- Practical - Ordered by what you need to build first.
Spec Structure
coding-agent-spec/
│
├── 00-SPEC-OVERVIEW.md # This file
│
├── foundations/ # UNDERSTAND (read first)
│ ├── 01-what-is-a-coding-agent.md # Mental model
│ └── 02-architecture-overview.md # Components and data flow
│
├── implementation/ # BUILD (core agent)
│ ├── 03-agent-loop.md # State machine spec
│ ├── 04-tool-system.md # Tool registration + execution
│ ├── 05-core-tools.md # Read/edit_file/create_file/Bash/glob/Grep
│ ├── 06-context-window.md # Token management
│ ├── 07-streaming.md # SSE/delta handling
│ ├── 08-verification.md # Course correction
│ ├── 09-memory.md # AGENTS.md, rules
│ ├── 10-minimal-agent.md # COMPLETE WORKING SPEC
│ └── 11-threads.md # Persistent conversation state
│
├── enhancements/ # LEVEL UP (production features)
│ ├── 11-subagents.md # Task delegation
│ ├── 12-course-correction.md # Meta-agent monitoring
│ ├── 13-multi-model.md # Model routing
│ ├── 14-advanced-tools.md # Kraken, look_at, etc.
│ ├── 15-permissions.md # Trust and safety
│ └── 16-skills.md # Extensible domain knowledge
│
├── reference/ # COPY-PASTE (implementation artifacts)
│ ├── tools/
│ │ └── TOOLS.md # Complete tool catalog + schemas
│ ├── prompts/
│ │ └── PROMPTS.md # Complete prompts (verbatim)
│ └── constants.md # All magic numbers
│
└── context/ # BACKGROUND (optional reading)
├── why-coding-agents-matter.md # Why this space matters
├── agi-thesis-deep-analysis.md # First-principles analysis
├── industry-comparison.md # How others do it
├── evidence-sources.md # Where this came from
└── open-questions.md # What we don't know
Reading Order
To Build a Minimal Agent (fastest path)
01 → 02 → 03 → 04 → 05 → 10
Skip to 10-minimal-agent.md for the complete spec once you understand the concepts.
To Build a Production Agent
01 → 02 → 03 → 04 → 05 → 06 → 07 → 08 → 09 → 10
To Understand Design Decisions
01 → 02 → context/why-coding-agents-matter.md → context/industry-comparison.md → context/agi-thesis-deep-analysis.md
What "Reproduction-Ready" Means
| Component | What You Get |
|---|---|
| Agent loop | State machine diagram + pseudocode + all transitions |
| Tools | Full JSON schemas, ready for tool_use API |
| Prompts | Complete text, not summaries |
| Constants | Every magic number with explanation |
| Decisions | Explicit choice points with tradeoffs |
After reading 10-minimal-agent.md, you should be able to implement a working coding agent in your language of choice.
Evidence Base
This spec is derived from reverse-engineering Amp Code (Sourcegraph's AI coding agent):
| Source | Version | What It Provided |
|---|---|---|
| Production bundle | v0.0.1769212917 (Jan 2026) | Current architecture |
| Source maps | v0.0.1744408301-dev | Original TypeScript (Rosetta Stone) |
| Changelog analysis | Apr-Dec 2025 | Evolution patterns |
The patterns here are validated against production code, not theoretical designs.
Quick Start
Want to build right now?
- Read
foundations/02-architecture-overview.md - Read
implementation/10-minimal-agent.md - Copy schemas from
reference/tools/ - Copy prompts from
reference/prompts/ - Implement in your language
Status
| Section | Status | Notes |
|---|---|---|
| Foundations | ✅ Complete | 2 docs |
| Implementation | ✅ Complete | 9 teaching docs + 7 spec docs |
| Enhancements | ✅ Complete | 6 docs (advanced topics) |
| Reference | ✅ Complete | Tools, prompts, constants |
| Context | ✅ Complete | 5 docs |
Spec started: 2026-01-24 Based on: Amp Code v0.0.1769212917 Methodology: Production code analysis