Enforceable goal constraints — delegation scope, spend limits, and time caps for autonomous runs

AutoGPT agents set goals, create tasks, and execute them autonomously. The agent decides what to do, does it, evaluates the result, and continues. The human sets the goal and walks away. Everything between the goal and the result is unsupervised.

This is the exact scenario where governance matters most. An unsupervised agent with access to the internet, filesystem, and code execution can:

• Browse to any URL, including authenticated sessions
• Execute arbitrary shell commands
• Read and modify any file the process has access to
• Spend unlimited API credits in a reasoning loop

The goal description is text. It's not an enforceable constraint. "Research market trends" and "buy everything on Amazon" look the same to the execution engine — they're both tool calls.

Delegation turns the goal into enforceable boundaries:

from agent_passport_system import create_delegation, govern_action

# Human sets goal + constraints (not just goal)
task_delegation = create_delegation(
    delegated_to=agent_key,
    delegated_by=human_key,
    scope=[
        "web:read",
        "file:read:/workspace",
        "file:write:/workspace/output"
    ],
    # no shell:execute, no web:post, no file:write outside /workspace/output
    spend_limit=10000,  # $100 API budget for the entire run
    expires_in_seconds=14400,  # 4 hour max runtime
    max_depth=0  # can't spawn sub-agents
)

# Agent autonomously decides to execute a shell command → blocked
result = govern_action(
    action={"type": "shell:execute", "command": "curl https://api.stripe.com/charges"},
    delegation=task_delegation,
    passport=agent_passport
)
# Blocked: shell:execute not in scope. Signed receipt.

The spend limit is the dead man's switch. When the API budget runs out, the agent loses all authority. No infinite reasoning loops burning $500 in API calls.

The time cap is the second safety net. After 4 hours, the delegation expires. The agent stops having authority to do anything, regardless of task completion.

Every action the agent takes produces an Ed25519-signed receipt. When the autonomous run finishes, the human has a complete, tamper-evident record of what the agent did — not what the agent says it did.

pip install agent-passport-system (v0.8.0, Apache-2.0) or npm install agent-passport-system (v1.36.2). Gateway at aeoess.com/portal.html.