AgentUse

An agent that needs two tools from a large MCP server loads all of them. A blocklist leaves gaps, and tools the server adds later load too. AgentUse 0.24 adds allowedTools: list names or wildcards like read_*, and every other tool on that server stays off, even new ones.

Cancellations usually show up in the data first: logins drop, a paid plan sits unused, a ticket asks about refunds. An AgentUse agent can check every Monday at 8am: read your accounts export, rank quiet accounts by MRR at risk, and draft one save play for each.

A rate limit, 5xx, or expired login at one provider can fail every scheduled agent on it. An AgentUse model alias can list fallback models across providers. A run that can't start on one moves to the next before any output or tool call, and its log names the model that ran.

When a scheduled agent fails, the recovery path shouldn't hide in logs. AgentUse keeps last-run health, cadence, and Run controls together on the fleet page, so an operator can spot the broken job, inspect its session, and rerun it without rebuilding the workflow.

AgentUse keeps orchestration visible instead of burying it in code. An intake agent routes work to activation, concierge, quotation, strategy, supplier, or triage. Each specialist is a readable agent file and can run directly from the operations UI.

An agent fleet is easier to run when it mirrors the work. The Kettlebase fleet groups 19 agents across six projects, from support triage to finance, on one serve screen. Last-run status and schedules stay visible for a fleet scan before opening a session.

Scheduled work is easy to define and hard to monitor. AgentUse v0.21 adds a 24-hour load strip and recent run health to Schedules. Every row can Run or Pause; a failed run adds Open session and Retry now, so operators can act without leaving the page.

Need a model provider AgentUse does not bundle? Provider plugins now install globally or per project, with local packages linked in place for development. Their models become available in the dashboard when you create or revise an agent.

Delegated agents should not disappear into separate tabs while they work. AgentUse v0.20 keeps each child run inside the parent session, including nested descendants and their live current activity, so you can follow a multi-agent workflow from one timeline.

Finding the run that needs you should not mean clicking through the dashboard. AgentUse v0.20 turns Ctrl/⌘K into one search across pending approvals, recent sessions, agents, pages, and quick actions. Open it anywhere and jump straight to the result.

When an agent run fails, the last thing you need is a scroll through every tool call to find out why. AgentUse v0.20 puts the failure and run timing in the Result card, then lets you filter the session log to agent activity, tools, or errors.

Connecting a model should not send you out of the dashboard. AgentUse v0.19.2 puts Anthropic, OpenAI, OpenRouter, OpenCode Go, and custom endpoints in Settings. Custom endpoints discover their models and verify the protocol with a real completion before saving.

Need to stop a scheduled agent during maintenance? AgentUse v0.19.2 lets you pause it from the dashboard or CLI without editing the agent file. The declared cadence stays visible, and pause state applies only to that deployment.

Not every review comment should rewrite how an agent behaves forever. Approval comments now revise only the current run by default. Check “Learn from this comment” when the correction should carry into future runs.

Pin an exact model version in an agent file and every new release means editing every file that named it. Model aliases remove that chore: name the line, not the version, like model: anthropic:claude-sonnet, and the agent tracks the newest release. A pinned version still wins.

A run that executes cleanly but has nothing to act on still gets logged as a win by default. report_incomplete(reason) gives agents a real third outcome: a distinct INCOMPLETE status with the reason attached, so a quietly stalled agent stops hiding behind a green checkmark.

What do you actually need to see scanning a fourteen-agent fleet at 8am? Not always Schedule, not always Last run. The serve agents view turns every column into a removable pill: drop the noise, keep what you're scanning for, and Monday's view opens already trimmed.

An agent breaks at 2am. Fixing it means opening the .agentuse file, tracing YAML by eye, then retyping context into a coding agent by hand. Open the agent in the serve hub, hit Send to Coding Agent, and paste the ready-made prompt, source and session id included.

Five drafted replies sit next to the five tickets they answer, wording locked the moment they're shown. await_human now snapshots referenced files at ask time, content-addressed. Name the file in the gate: approval holds even if it changes after.