Skip to content

beogip/code-first-agents

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Code-First Agents

Design patterns for building reliable AI agents by moving deterministic work from the LLM to code.

Your agent works. Until it doesn't. And you can't tell why.

The LLM picks the wrong branch. It skips a step. It hallucinates a field name. You can't write a test for any of it because the decision happened inside a black box.

Code-First Agents is a set of patterns for moving decision-making out of the LLM and into deterministic code. The LLM orchestrates. Code executes. The more you push into code, the less you depend on probabilistic output for things that should be deterministic.

Patterns

Pattern Description
Deterministic Tools How to build CLI tools that do deterministic work and output JSON. The tool contract, the output spectrum, and progressive examples.
Skill Orchestration How to write SKILL.md files that consume tool output. The LLM as executor, chaining tools, guards, and the verbatim execution principle.

For the formal specification of both patterns, including the contract, trade-offs, reference implementations, and evolution history, see specs/code-first-agents.md.

The spectrum

Tools exist on a spectrum based on how much decision-making they absorb from the LLM:

Level Tool returns LLM does Example
1. Data Structured facts Interprets, decides get-issue-signals.ts
2. Classification Data + classification Reads the class, branches classify-issue.ts
3. Instructions Literal procedure Executes verbatim analyze-issue.ts

At Level 3, the tool becomes a prompt factory. The LLM is a pure executor.

Examples

The examples/ directory contains a complete GitHub issue planning workflow that demonstrates the three levels. All tools run with Bun:

# Level 1: raw signals from the issue body
bun examples/tools/get-issue-signals.ts --owner acme --repo app --issue 42

# Level 2: deterministic classification
bun examples/tools/classify-issue.ts --owner acme --repo app --issue 42

# Level 3: complete planning procedure
bun examples/tools/analyze-issue.ts --owner acme --repo app --issue 42

See examples/skills/plan-issue/SKILL.md for a full skill that orchestrates analyze-issue.ts at Level 3.

Site

The full documentation lives at https://beogip.github.io/code-first-agents/.

The site source (HTML, CSS, markdown mirrors) lives in the gh-pages branch. This branch (main) holds the example tools, the example skill, and the repo metadata.

Contributing

See CONTRIBUTING.md for guidelines on proposing new patterns and adding examples.

License

MIT

Author

Built by Juan Gipponi.

About

Design patterns for building reliable AI agents by moving deterministic work from the LLM to code

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors