Lead Agent · reference lead agent / orchestrator

Give it a goal. Every step is governed.

The lead agent plans, starts sub-agents with passports bound to it, and sends every tool call through a guard. You can suspend, reinstate or revoke any agent, run a shutdown test, and verify the audit ledger yourself.

Live in this tab · deterministic model · no account, no key, nothing sent

Conversation

Mock model · this tab

Lead agent

Tell me what you want done. I will plan it, start governed sub-agents, and show you each guard decision. Try a suggestion below.

Fleet

Depth cap 3

Starting the lead agent…

Suspends the whole fleet, checks nothing can act, then reinstates it.

Guard decisions

0 tool calls
  1. Each tool call appears here with its Action Level and decision.

Audit ledger

SHA-256 hash chain
Records
0
Head
none yet

Every record holds the hash of the one before it. Change any byte and verification fails.

Offline check: lead-agent verify-audit lead-agent-audit.json verifies an export from this page with the same rules.

How the guard decides

One order, every call.

  1. Level 5 is denied. Spending money or deleting records is never an agent action here, whoever asks.
  2. Passport and lineage. The agent and every parent above it must be active. Suspending or revoking a parent stops all its descendants.
  3. Inventory and ceiling. The tool must be in the agent's passport, and its Action Level at or under the passport ceiling.
  4. Agent Fleet Controls. Spawning, network calls and spending must be granted to this agent, and no spawn may exceed delegation depth 3.
  5. Level 4 approver. The operator's automated approver clears listed Level 4 actions. Anything else at Level 4 waits for a person.
  6. Levels 1 to 3 run unattended. Every decision, result and control lands in the audit ledger.

Optional

Use your own models.

Run the same lead agent on your machine. It uses the same governance code with signed passports, and calls OpenAI, Anthropic or xAI only when you set that provider's key in its environment.

pip install "agentic-ai-governance @ git+https://github.com/GLOBAL-AI-GOVERNANCE/agentic-ai-governance@main"
export OPENAI_API_KEY=…   # or ANTHROPIC_API_KEY, or XAI_API_KEY
lead-agent serve --port 8765

Not connected. The page runs on its own until you connect.

Keys stay out of the browser. This page never asks for, stores or sends an API key. Keys live only in the server's environment on your machine, and the server sends each one only to the provider you chose. The page talks only to the address above.

Boundaries

What this is, and is not.

A reference implementation and a demonstration. In this tab, passports are simulated and unsigned and tools are simulated; nothing leaves the page. The local server issues signed passports and validates them with the published tools. Results are non-authorizing (authority_effect: NONE): they do not certify an agent, grant permission, or replace an accountable person.

Source and tests · Works with · Start here