Documentation

Introduction

FirstHelm is the human-first control layer for autonomous AI agents. It gives organizations a single place to register their AI agents, define the goals those agents work toward, set the guardrails that keep them safe, and step in whenever a human decision is needed.

It is built for teams who are already running autonomous agents — research agents, coding agents, operations agents, multi-agent crews — and who have realized that giving software workers the ability to call APIs, send email, and execute code without oversight is a risk no compliance, security, or engineering leader can accept. FirstHelm does not replace your agent framework; it wraps it with the governance, approvals, audit trail, and real-time intervention that production deployments require.

The key concepts are simple: Agents are your autonomous workers; Missions are the goals you assign them; Constraints are the rules they must obey; Approvals are the moments a human must sign off; Interventions are the moments a human steers; and the Activity Log is the immutable record of all of it.

Getting Started

1. Create an account. Sign up at firsthelm.dev and verify your email. Your first organisation is created automatically.
2. Connect your first agent. Go to Agents → Register. Provide a name, framework, type, the agent's endpoint URL, and its capabilities. FirstHelm will mark the agent idle until it reports in.
3. Define a mission. Go to Missions → New Mission. Give it a title, a goal, success criteria, a priority, and a budget limit. Assign one or more agents.
4. Set your first constraint. Go to Constraints → Add. Start with an approval gate scoped to the mission, requiring human sign-off for any high-risk action. Test it against a sample payload.
5. Respond to your first approval. When your agent attempts a gated action, an approval request appears in the Approvals queue and the notification bell. Review the proposed payload, then approve, reject, or edit-and-approve. Your decision is logged with your identity and timestamp.

Core Concepts

Agents

Autonomous AI workers registered to the platform. Each agent has a framework (LangChain, CrewAI, AutoGen, OpenAI, Anthropic, Vertex, or custom), a type, a status, an autonomy level, capabilities, and an endpoint it can be reached at.

Missions

High-level goals you assign to agents. A mission carries a goal, success criteria, priority, budget, and progress. Missions move through draft → active → paused → completed / failed / aborted.

Tasks

The steps within a mission, assigned to specific agents. Tasks track their own status (pending, active, completed, failed, blocked) and results.

Constraints

The guardrails that keep agents safe: budget_limit, action_restriction, approval_gate, rate_limit, time_window, and forbidden_action. Each is scoped globally, to a mission, or to an agent, and carries a severity: advisory, warning, or blocking.

Approvals

Human sign-off gates. When an agent proposes an action above the risk threshold, execution pauses until a human approves, rejects, or edits the request — or it times out.

Interventions

Direct human steering mid-flight: pause, resume, redirect, approve, reject, edit, kill, or rewind an agent or mission. Every intervention is recorded with the actor and reason.

Autonomy Tiers

A 0–100 score representing how much trust an agent has earned. New agents start at 0 (every significant action needs approval). Successful history raises the tier; failures, violations, and interventions lower it. Higher tiers mean fewer approval interruptions — but high-risk actions always require sign-off.

Using FirstHelm

Each page in the app serves a distinct purpose:

Integrations

FirstHelm connects to your agents wherever they run. Supported frameworks:

• LangChain — register the agent with its framework set to langchain and expose its run endpoint.
• CrewAI — connect a crew as a multi_agent type; FirstHelm treats the crew as a single managed unit.
• AutoGen — register the conversation manager; multi-agent status is reflected automatically.
• OpenAI Agents — framework openai; point the endpoint at your hosted agent runner.
• Anthropic agents — framework anthropic; connect the agent's endpoint and capabilities.
• Custom agents — framework custom; any HTTP or webhook endpoint that conforms to the FirstHelm activity schema.

FAQ