AgentUse v0.18.0: smaller agents and calmer operations

AgentUse v0.18.0 is on npm. This release makes agents simpler to create and easier to operate: the official creator guidance now produces smaller files, and agentuse serve has a calmer, denser interface for watching real work.

npm install -g agentuse@latest

Start with the job, not an agent framework

Agent files tend to grow before they run. A short request becomes a long prompt with speculative branches, copied configuration, and instructions for decisions a capable model could already make. The extra text looks thorough, but it gives the agent more rules to reconcile on every turn and more ways for those rules to contradict the job.

The version-matched creator skill now starts from the smallest execution contract that can do the work:

---
model: anthropic:claude-sonnet
description: "Summarize support trends for the product team"
---

## Task
Read this week's support conversations and identify the recurring problems.

## Output
Write a concise brief with evidence, impact, and the next decision to make.

Frontmatter is added only when the workflow uses it. No schedule until the job actually recurs, no sub-agent roster before distinct roles are necessary, no reasoning budget for mechanical work, and no approval protocol copied into the body when the runtime already supplies it. The same test applies to prose: if removing a sentence does not change what a good agent would do, it should probably go.

The target for an ordinary agent is now a focused 300–700 word body. That is a starting discipline, not a hard limit: a judgment-heavy agent can be shorter when the objective and inputs carry the context, while a workflow with real ordering constraints should still state them. The point is that complexity has to earn its place.

Fetch the guidance that matches the AgentUse version on your machine with:

agentuse skills get creator

The operations UI gets out of the work's way

agentuse serve has been redesigned across Home, Approvals, Sessions, Agents, Schedules, and Stores. The new visual system is quieter and denser, with responsive layouts and session logs that put machine metadata in the background and the work itself up front.

Approval queues received the most practical attention. Pending gates are grouped by agent, groups are ordered by their newest request, and the newest work appears first inside each group. Pick-one approval cards lead with risk and fold long drafts and commands until you need to inspect them. Once you decide a gate, stale live data can no longer make it briefly reappear.

The timeline is clearer after a reviewer comments on delegated work too. A parent session now records the comment as a timestamped revision request, labels an active child revising, distinguishes a reopened awaiting-approval gate, and identifies Judge attempts as automated pre-review. Long-running sessions also stay visible after their creation date falls outside the selected activity window; activity now follows the session's latest update.

These are small distinctions when you run one agent once. They become the interface when several agents are working, approvals arrive throughout the day, and the session record is how you reconstruct why a result changed.

Better performance across AgentUse

The v0.18.0 performance pass was measured against a real AgentUse installation with 3.9GB of session logs, 4,151 session files, and seven served projects containing 213,000 files:

Session listing is faster because AgentUse prunes irrelevant directories and reads session metadata concurrently. The file watcher no longer walks every project with a general-purpose glob every 15 seconds; a quiet project now backs off to a 60-second scan. Parsed stores and search material are cached, and long conversations no longer recount their entire history on every step. The dashboard's Brotli-compressed critical path also fell from 55.2KB to 41.3KB.

MCP discovery now starts one server process instead of two, authentication no longer adds a 138–152ms file-lock stall to concurrent model requests, and idle workers are actually recycled after memory-heavy runs. Response caches are bounded, skill discovery is memoized, and long-lived processes do fewer whole-project scans.

There are reliability changes in the same layer:

• Outcome recovery now receives the complete tool trace. If a model finishes without declaring complete or incomplete, the recovery turn can report the outcome without repeating side effects or inventing a blocker.
• Sub-agents now use the model declared in their own file instead of silently inheriting the parent's configured model. Explicit command-line model overrides still cascade through the full agent tree.
• Failed sub-agent runs name the concrete model and provider that failed, including authentication errors, so a fallback problem is diagnosable from the parent session.
• Persistent store content now comes with an explicit trust boundary. Stored prose cannot grant itself authority, and time-sensitive claims about credentials, quotas, networks, locks, providers, or services must be checked again before an agent uses them to skip or block work.
• Resume and preflight state changes are more defensive: an already-executing decision cannot reopen, a failed preflight restores its approval lease atomically, and PID reuse is rechecked before a live run is declared orphaned.

v0.18.0 is less about adding another layer to AgentUse than removing friction between the layers already there. Agent files say only what the job needs. And when the agents are running, the operational surface makes the next decision easier to see.