Terminal Agent vs. Web Chat: Pick by Blast Radius
Every major AI lab now ships a command-line agent that runs on your machine with access to your files and shell. The deciding factor is which tasks you deliberately keep out of it. The deciding factor is simpler than most teams make it. The Idea: Scope the Damage, Not the Capability The better question when evaluating a terminal agent is blast radius: if this thing does exactly the wrong thing,…
When comparing terminal agents to web chat interfaces, the primary consideration should be their impact radius. A web chat window's blast radius is essentially nonexistent, with the worst-case scenario being the inadvertent copy-paste of incorrect text. Conversely, CLI agents with write access to a git repository pose a higher risk, potentially altering uncommitted changes. An agent possessing production credentials carries the highest blast radius, threatening the entire business.
The blast radius is a characteristic of the task at hand rather than the underlying model. By assessing this metric during planning meetings, teams can determine the appropriate level of autonomy before code is written. Certain tasks, like refactoring tests in a temporary branch, present a low blast radius, allowing the agent to operate with greater autonomy.
Conversely, editing infrastructure configurations that trigger deployments upon merge requires a different decision-making process despite using the same underlying model and prompt quality.
A practical implementation of this approach involves a three-tiered system with approval modes and sandboxing. Tier 1 focuses on scratch work, where the agent runs autonomously with git serving as the undo mechanism. Tier 2 involves real repositories, where the agent proposes diffs for human approval before execution. Tier 3 encompasses tasks involving live credentials, demanding a read-only approach with no shell access.
The critical aspect is not the flag setting but the decision-making process that determines the tier. Three essential questions guide this process: reversibility within five minutes, presence of live secrets in the environment, and the visibility of errors. If the answers are "no" to reversibility, credential exposure, and error visibility, the task may not be suitable for autonomous execution.
By applying these three factors—reversibility, exposure of live secrets, and error visibility—any project manager or engineer can evaluate AI agent tasks before writing a single line of configuration. Approval modes and sandboxing serve as safeguards, enabling rapid progress on low-radius tasks without jeopardizing the remaining 20%. It is worth reflecting on which tasks have already been transitioned into auto-approve loops and any instances where moving back to manual review proved beneficial.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
