Chat Polish — Task Budget and Adaptive Thinking
May 1, 2026 · InfraScout Team
A long playbook run can fan out into dozens of tool calls before it returns a finding, and until now the chat gave you no real sense of how much of the AI usage budget that single task was burning. This update adds two quiet but high-value improvements to the chat experience: a live task budget that tracks what the current task has consumed, and an adaptive thinking display that only surfaces Claude's reasoning when it actually used extended thinking for a step.
A Live Task Budget for Every Chat Task
When you ask Claude to run a multi-step assessment — enumerate Entra ID conditional access policies, walk a Defender incident, audit an Azure subscription — a single message from you can trigger a long sequence of tool calls before you see a final answer. Each iteration consumes tokens against your AI usage budget, and previously the only feedback came after the fact.
The chat now tracks a per-task budget that starts fresh at the beginning of each turn and ticks down as Claude works through the steps. You can watch the remaining budget shrink in real time, so a long-running playbook never quietly runs over. Just as importantly, Claude itself sees the same budget and can pace its own work — wrapping up cleanly, summarizing what it has found, and handing back control before the budget runs out, instead of being cut off mid-step.
The practical effect is more predictable behavior on the assessments that matter most. A short factual question still completes in a single quick turn. A deep playbook run takes longer, but it stays within bounds and produces a coherent finding rather than a truncated transcript.
TIP
The task budget resets at the start of each chat turn — it bounds a single task, not the whole conversation. Long conversations with many distinct asks remain unaffected.
Adaptive Thinking — Only When It Matters
The chat used to render Claude's extended thinking as a fixed, always-on panel. That was useful when reasoning was happening, but distracting when it wasn't — quick lookups got the same heavy UI as deep analysis, and the signal got lost in the noise.
The display is now adaptive. When Claude uses extended thinking on a step, you see a concise summary of that reasoning inline with the response. When a step doesn't need it — listing connected agents, fetching a known piece of configuration, answering a one-liner — the thinking UI stays out of the way. The result is a quieter chat for short Q&A and a clearer signal in the moments when reasoning actually matters.
For audit-heavy work, this also makes it much easier to scan a conversation after the fact. Reasoning appears next to the steps that called for it, so you can see at a glance which findings were the product of extended analysis and which were straightforward.
More Accurate Token Accounting
Alongside these visible changes, this release also tightens how AI usage is counted during a long task. Multi-step playbook runs now consume noticeably fewer tokens for the same work, which translates directly into a smaller, more predictable spend per task — and a more honest reading on the task budget you see in the UI.
Pairs Well With the Executions Audit
The task budget is a live signal in chat, but it isn't the only place the data lives. The Audit & Compliance Executions view records every chat task and tool invocation that runs through the server, so administrators can review budget usage, tool calls, and outcomes long after a conversation has ended. Together, the live budget and the executions audit give you both real-time guardrails and an after-the-fact paper trail for AI-driven assessments.
Try It
The new chat experience is live for everyone running the latest InfraScout server. Open a conversation, kick off a playbook, and watch the task budget tick down as Claude works. If you have feedback on what the chat should show next — or what other signals would help your team trust long-running AI assessments — write to us at info@infrascout.cloud.