How it works

01 Create a Phaser

3D view unavailable. Use the scene and component controls to explore the flat diagram.

Task: troubleshoot application crash in prod/app

Illustrative review flow. Alice assigns a crash investigation to an agent with zero standing privileges. A model approves a read-only log plan; Alice denies a production restart: Too risky, find other means. No executor starts for the denied plan. Approval returns to the requesting agent before it dispatches the approved plan to a separate execution beam. Only the executor connects to Kubernetes, through the Protocol-Aware Proxy in the virtual network. Results return to the requester; its privileges never change. Commands and approvals are simulated. A restart is not a guaranteed fix.

Illustrative creation sequence: Alice delegates the zero-standing-privileges role to the Phaser. The role is inherited only when the delegation arrives; it does not grant standing resource access. No commands are executed and no cloud resources are created. Each scene plays once and holds. Scene 2 is an exploded logical view of the same Phaser; the container runs inside the VM, not above a separate machine. Use the component controls to inspect each part. Use Play to continue or watch a completed scene again. Scene 3 illustrates a retained Phaser environment across 12-hour beam rotations. A fresh base image replaces the micro-VM, not the /home checkpoint. A fork copies /home into a separate Phaser with its own runtime; private keys are not copied. Build commands and revise/evaluate cycles are illustrative: no commands run, and snapshot restores do not guarantee model improvement. Scene 4 shows illustrative AWS account and resource discovery, IAM enrollment, scoped authorization and concurrent audit capture. The zero-standing-privileges role does not itself grant resource access: the illustrated policy permits pod reads in prod/app and model inference before requests are sent. Astra and GLM 5 are illustrative model labels, not verified Bedrock model IDs or availability claims. Kubernetes and database traffic use the Protocol-Aware Proxy; only model inference uses the LLM proxy. Both proxies are inside the virtual network. Risk scores and summaries are simulated, not actual analysis.