We’re excited to announce Ag.dev’s Early Access: a cloud platform for building, testing, and scaling autonomous, headless AI agents—no code required.
Find out more about:
What Ag.dev is and why it’s different
A concrete “deep company research” example you can build in minutes
How to join our community and share feedback
What is Ag.dev?
Ag.dev is a development platform for “headless agents” — autonomous AI workers that run in the background, triggered by events or API calls, and return structured, production-ready results.
Instead of wiring complex orchestration logic, you describe:
The agent’s goal and constraints in natural language
The tools it’s allowed to use (web search, news, browsing, or your own via MCP)
The input schema it accepts and the output format it must produce
Ag.dev handles execution, observability (full run traces), and scaling (API + managed infrastructure).
Why teams use Ag.dev:
Automation at the speed of prompt: No code required to define, test, and iterate.
Full observability: See planning, reasoning, and tool calls for every run.
Scales by design: One-click deploy; invoke via API/webhooks; batch across thousands of inputs.
Extensible: Connect private systems via MCP for secure, high-trust automation.
Predictable outputs: Return documents or schema-validated JSON.
Getting Started: Build a “Deep Research Profile” Agent
In this example, you’ll create a headless agent that produces a detailed, narrative research brief for a target company. The output will be a document (not JSON).
Get started:
Sign up: https://auth.ag.dev/signup
Quickstart: https://docs.ag.dev/quickstart
Step 1 — Create an Agent
Configure the agent with:
Output format: Document (text)
Built-in tools: Enable Web Search, News, and Browse Website
Model: Choose a model suited to research tasks (you can adjust later)
Goal Prompt (templated with Handlebars):
You are a meticulous market research analyst.
Task:
Create a deep research profile for {{ company }}
Instructions:
- Prioritize authoritative sources and recent information (past 12–18 months).
- Cross-check claims; avoid speculation. If data is unavailable, note the gap.
Deliverable (Document):
Include clear section headings and concise, source-backed content:
1. Company Overview
2. Key Persons
3. Latest News
Constraints:
- Be objective and succinct.
- If browsing reveals conflicting data, explain the discrepancy and favor more recent primary sources.
Step 2 — Define the Input Schema
Use JSON Schema to specify the inputs your agent expects. For this example, the only required field is company, with optional fields for disambiguation.
{
"type": "object",
"required": ["company"],
"properties": {
"company": { "type": "string", "description": "Target company name" }
}
}Step 3 — Configure Tools and Settings
Select a model stack for your Agent, for example “openai/reasoning-low”. This model stack uses reasoning models with low reasoning time to create and update the agent's plan.
In the advanced settings section, enable the context compression. This option compresses the context of your agent while running on a regular basis and enables long-running and complex tasks to run without hitting a context lenght limit.
Step 4 — Test Your Agent
Provide a sample input, for example:
{ "company": "ACME Inc."}
Start a test run and watch the live event stream:
reasoning: intermediate thinking as it evaluates sources
tool_call: search, news, and browse actions with parameters
Refine quickly:
Tighten or expand the instructions
Emphasize recency or specific data ranges
Adjust headings or tone to match your brand
Tip: If you see noisy sources in the trace, add directives like “favor primary sources (company site, filings, reputable outlets); deprioritize low-authority blogs.”
Step 5 — Deploy and Integrate
Save your agent to deploy to enable production runs.
Invoke via API/webhooks from your product, CRM, data pipelines, or internal tools.
Batch runs: feed a list of companies to generate many profiles in parallel.
Minimal invocation pattern (illustrative):
# Create a run
curl -X POST "https://api.ag.dev/agents/{AGENT_ID}/runs" \
-H "Authorization: Bearer YOUR_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"input": { "company": "ACME Inc."}
}'
# Poll run status or subscribe to events, then fetch the final document resultThe run result will contain your document output plus a list of sources if browsing was used.
Step 6 — Extend with Your Tools (Optional)
Connect internal systems via MCP (e.g., CRM, data warehouse, procurement, ticketing).
Expose only the functions your agent needs and keep strong boundaries.
Re-test with observability to validate behavior before scaling.
What You Can Automate Today
Sales and market research: Generate deep company or account briefs on demand.
Data enrichment and analysis: Summaries, watchlists, and change detection.
Personalization at scale: Tailored onboarding or outreach content based on fresh research.
Content generation: Reports, memos, and structured documents ready for stakeholders.
Who the Ag.dev Early Access Is For
Builders in operations, marketing, sales, product, and data who need event-driven, repeatable automation—not another chat UI.
Teams that require dependable outputs and transparent audit trails.
Organizations planning to connect private data and APIs in a secure, observable way.
What to expect in the early access:
A stable core with rapid iteration on ergonomics, observability, and integrations.
Help from our team to ship your first agent to production.
Early access features; we’ll communicate any breaking changes in advance.
Notes: The early access version is hosted in the US and we are currently logging all interactions, agents and agent runs in full text to improve our platform. The platform is provided as is and is subject to change and further improvements.
Join the Community: Need Help? Want Feedback?
Sign up: https://auth.ag.dev/signup
Quickstart: https://docs.ag.dev/quickstart
Join the Ag Discord: https://discord.gg/cPSJbuKDse
Tell us what you’re building, share your prompts, and get rapid help from the team and other builders. We’re especially eager for feedback on:
Goal Prompt ergonomics and document formatting controls
Tooling gaps (built-in vs. MCP integrations)
Which traces/logs make you fastest to production
Batch and webhook patterns you rely on
Let’s build reliable, headless automation—at the speed of prompt.