Building an Exsecutor Agent

Welcome to the Builder SDK. Exsecutors are purely declarative aggregations of logic (Skills) and capabilities (Tools). As a builder, you do not write python scripts for the agent's brain—you author its Markdown skills and its YAML manifest, and let our trusted Ralph Loop execute it.

1. Initialization

Install the Exsecute CLI and initialize a new agent repository.

npm install -g @exsecute/cli
exsecute login
exsecute init agent my-awesome-agent
                    

This will scaffold the required directory structure:

my-awesome-agent/
├── exsecutor.yaml            # Main configuration
├── persona/
│   └── system_prompt.md      # The agent's core identity and constraints
├── skills/
│   └── default_skill.md      # Domain heuristics
└── tests/
    └── mocks.json
                    

2. Defining Capabilities

Open `exsecutor.yaml`. This file defines your agent's required inputs, schedule, capabilities, and permissions. You can use verified marketplace tools or reference your own private tools.

version: "1.0"
agent:
  name: "my-awesome-agent"
  objective: "Analyze AWS VPC for public exposure."
  
  allowed_tools:
    - name: "exsecute/aws-config-fetcher"
      version: "^2.0.0"
    
  permissions:
    allowed_secret_keys:
      - "aws"
                    

3. Local Sandbox Testing

Before submitting your agent to the marketplace validation pipeline, you can run a local mock of the Ralph Loop using your own OpenAI or Anthropic key.

export EXSECUTE_LOCAL_LLM_KEY=sk-ant-...

exsecute run --local \
  --param aws_account_id=123456 \
  --secret aws=AKIA...
                    