Agentform™

Define AI agent systems declaratively using Agentform™ native schema

Think Infrastructure as Code, but for AI agents

Why Agentform™?

Most AI agent frameworks require you to write imperative code - managing state, handling retries, wiring up tools. Agentform takes a different approach: describe your agents declaratively in Agentform native schema, and let the runtime engine handle the rest.

1
2
3
4
5
6

agent "reviewer" {
  model        = model.gpt4o
  instructions = "Review code for security issues"
  allow        = [capability.read_file, capability.get_diff]
  policy       = policy.strict
}

The result: Your agent configurations become version-controlled artifacts that are easy to review, share, and reproduce. The native .agentform format provides type safety, explicit references, and improved editor support.

Quick Start

1. Installation

1

pip install agentform-cli

2. Set up your API key

1

export OPENAI_API_KEY="your-openai-key"

3. Create an agent spec

Create a file called my-agent.agentform:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

agentform {
  version = "0.1"
  project = "my-first-agent"
}

variable "openai_api_key" {
  type        = string
  description = "OpenAI API key"
  sensitive   = true
}

provider "llm.openai" "default" {
  api_key = var.openai_api_key
  default_params {
    temperature = 0.7
    max_tokens  = 2000
  }
}

policy "default" {
  budgets { max_cost_usd_per_run = 0.50 }
  budgets { timeout_seconds = 60 }
}

model "gpt4o_mini" {
  provider = provider.llm.openai.default
  id       = "gpt-4o-mini"
}

model "gpt4o" {
  provider = provider.llm.openai.default
  id       = "gpt-4o"
}

agent "assistant" {
  model           = model.gpt4o_mini
  fallback_models = [model.gpt4o]

  instructions = "You are a helpful assistant. Answer questions clearly and concisely."

  policy = policy.default
}

workflow "ask" {
  entry = step.process

  step "process" {
    type  = "llm"
    agent = agent.assistant

    input { question = input.question }

    output "answer" { from = result.text }

    next = step.end
  }

  step "end" { type = "end" }
}

4. Run it

1
2
3
4
5

# Validate your spec
agentform validate my-agent.agentform

# Run with input
agentform run ask --spec my-agent.agentform --input '{"question": "What is the capital of France?"}'

Features

Feature Description
Native Schema Define agents, workflows, and policies in type-safe .agentform format with explicit references
Modules Terraform-style reusable modules for sharing agent configurations via Git
Multi-Provider Use OpenAI, Anthropic, or other LLM providers interchangeably
Multi-Agent Coordinate multiple specialized agents with conditional routing
MCP Integration Connect to external tools via Model Context Protocol servers
Policy Enforcement Set budgets, timeouts, and capability limits per agent
Human-in-the-Loop Built-in approval gates for sensitive operations
Execution Tracing Full visibility into workflow execution for debugging

Next Steps

Built with ❤️ for the AI agent community

Built with ❤️ for the AI agent community

GitHub • PyPI