Auto Mode for Frontier Investigation Agents
Safe autonomy for cloud operations. On your terms.
Frontier Investigation Agents find the why and run the runbooks you pre-approved, inside your policies. Start with every action reviewed. Raise autonomy only when the evidence earns your trust.
Approving every action by hand is the bottleneck.
Manual approval keeps a human in the loop for every action. It's safe, but slow, and it sets your MTTR.
Start in Manual. Graduate to Auto.
Agents read freely in both modes: metrics, logs, dependencies. What changes is who says yes to a write.
Manual: the agent proposes, you decide
Every pod restart, scale event or config change waits for a thumbs-up. Each tool is set to always allow, needs approval or never allow.
Auto: pre-approved actions run
If a Runbook or Agent Policy covers it, the agent executes without interruption. Novel actions escalate to a human.
Trust grows with every Runbook
Every human approval is a Runbook candidate, so coverage grows from your real operational patterns.
Every write passes the gate.
The Action Classifier checks each proposed action against your policy and returns a verdict in under 200ms.
Allowed actions run
A Runbook or Agent Policy match runs at once. The agent then checks the outcome and confirms it worked.
Novel actions wait
Anything without a pre-approval is queued for a human, with the target, the parameters and the reason attached.
Hard limits never move
No auto-approval for deleting production databases, clusters or VPCs, for admin IAM grants, or for changes to more than 50 resources at once. These guardrails can't be overridden.
Questions, answered.
What is Auto Mode?
Auto Mode is CloudThinker's approval-gate layer for autonomous AI agents. Each Skill runs at one of three configured levels: Notify (post the recommendation, wait), Act-with-Approval (open a Merge Request with a scoped diff and rationale), or Autonomous (ship under defined guardrails). The level is set per environment and per service, encoded as policy rather than ceremony.
How is Auto Mode different from approving every action manually?
Manual approval keeps the human in the loop for every action. Safe but slow, and the bottleneck on MTTR. Auto Mode preserves manual approval where it matters (production, the auth service, novel Skills) while letting well-trodden paths (cost right-sizing, dependency upgrades, known-good runbooks) ship under guardrails. Teams typically start at Notify, promote per Skill to Act-with-Approval, then graduate to Autonomous on a Skill-by-Skill basis.
Can a misbehaving agent escalate itself past Auto Mode?
No. The policy is encoded outside the agent, the credential is brokered at task time at a scope the policy authorises, and execution runs inside an isolated sandbox. A compromised agent still cannot reach a production system without the team's policy saying yes to the specific action.
What does the Auto Mode audit trail look like?
Every action carries the request, inputs, tool calls, outputs, operator (human or agent), and the policy gate it cleared into a tamper-evident log. Post-incident reconstruction is one query, not Slack archaeology. The same audit substrate feeds long-term memory for Resolve.
Explore the platform.
- Platform OverviewOne unified layer under every agent.
- AI-DLCAI writes the code. Agents run it in production.
- Connections100+ secure connections give every agent the full picture.
- Self-hosted sandboxesCloudThinker, on your own machine.
- Skills FrameworkTurn operational knowledge into safe agent action.
- Security & ComplianceSecurity by design, at every layer.
Ready to graduate from manual approvals?
Start in Manual. Define your first Runbook. Watch your team sleep through the next 2AM incident.