VIVUU · Guides
Running AI agents with a human at the right step
Guides and field notes on running AI agents safely: control planes, human-in-the-loop approvals, guardrails, audit trails, monitoring, failure modes and costs.
Guides
How to put AI agents to work without handing them the keys: what a control plane is, which actions need a human, and what an approval must show to be worth anything.
- Definitions
What Is an AI Agent Control Plane?
A plain definition of the AI agent control plane: policy, approvals, evidence and audit between agents and consequences, and how it differs from observability.
- Agent operations
Monitor Claude Code from Your Phone
How to watch and approve Claude Code and other background coding agents from your phone: the state you need, today's workarounds, and a purpose-built terminal.
- Agent governance
Human-in-the-Loop Approvals for AI Agents
How human-in-the-loop approval works for AI agents: when an agent should stop and ask, what a reviewable approval contains, and policy tiers from ALLOW to DENY.
- Agent governance
AI Agent Audit Trails: Every Action Attributable
What an AI agent audit trail should record (actor, action, target, result, evidence), why append-only matters, and how denials and approvals appear in it.
- Agent governance
AI Agent Guardrails: The Four Layers That Matter
The four layers that actually constrain an AI agent (prompt, tool, policy and gate), why prompt-level rules are the weakest, and concrete examples of each.
- Agent governance
When to Require Approval for an AI Agent
A concrete test for which AI agent actions need human approval: reversibility, blast radius and leaving the sandbox, plus what the approval screen must show.
- Agent governance
Human in the Loop vs Human on the Loop for AI Agents
Human-in-the-loop vs human-on-the-loop vs human-in-command for AI agents: what each means for approvals, and why to choose per action rather than per system.
- Running agents
Monitoring AI Agents in Production: What Matters
What to monitor when AI agents run in production: signals that predict harm instead of describing activity, why token dashboards mislead, and watch vs control.
- Founder role pack
AI Chief of Staff: Decisions, Not Chat
What an AI chief of staff should do for a founder: watch revenue and customer signals, work the routine, and bring decisions with evidence, not a chat box.
Field notes
First-hand data from running coding agents against real repositories behind an approval gate. What broke, what it cost, and what we changed.
- Running agents
AI Agent Failure Modes: 6 Seen in Production
Six failure modes from running coding agents on real repos behind an approval gate: confident wrong plans, empty successes, vocabulary drift, lost workspaces.
- Running agents
When AI Agents Are Confidently Wrong: A Real Case
An agent proposed inverting a live content-security policy with a fluent, invented rationale. The summary read well; the diff was backwards. What changed after.
- Running agents
What AI Coding Agents Actually Cost Per Task
Measured cost of a gated coding agent end to end: about ten cents a task and two model calls, and why the model bill is cheaper than review time and re-runs.