1. Docs
  2. AI
  3. Autonomous Agents

Autonomous Agents

Autonomous agents are AI workers that look after your project on a rhythm: after each deploy, every morning, every week or when you ask. They read your project, write a report and record any change they want as a proposal that a person applies.

7 min read Updated CustomersAdmins

In short

On this page you can

  • Install ready-made agents from a catalog of fourteen, grouped by when they run.
  • Build a Custom agent with your own instructions and tools.
  • Run an agent on demand with Run now and read its report.
  • Review the changes an agent proposes and apply or dismiss each one.
  • Cap every agent with step, per-run cost and daily budget limits.

How to use it

  1. Open Agents in the main menu and pick your project in the header.
  2. Click Install new agent, choose an agent and click Install.
  3. Review the settings, save, then switch the agent on (it starts paused).
  4. Let it run on its trigger, or click Run now.
  5. Open the run from the card or the Runs tab and apply the proposals you agree with.

At a glance

Where
Main menu › Agents (page title Autonomous Agents). Tabs: Agents and Runs.
Who
Every role (Super Admin, Admin, User) in projects they can open. A User role’s agents can be refused tools its role has no permission for.
You need
A project selected in the header. Deploy agents also need a code repository; Deploy verification needs a Smoke pipeline.
Runs on
The SmartestQA server with the project’s AI connection, or a scheduled task in Claude Desktop.
Produces
A run with a summary, a report (also as PDF), every step it took, and proposals for changes.
Safety
New agents start paused. A server run never writes: every change waits for a person to apply it.
Project types
WEB, MOBILE and DESKTOP. Agents read runs, reports, defects and the app map, which exist in every project type.

Start from your question#

Where agents fit#

Agents come after the order of work, not inside it. Once you have tests and pipelines, agents keep watching them so nobody has to check every morning by hand.

  1. 1DiscoverPages, controls, rules
  2. 2AnalyzeRequirements
  3. 3AutomateTest cases
  4. 4SchedulePipelines
  5. 5ReportResults
  6. 6Maintenance + AgentsWatch, triage, report

The last stage of the Setup Roadmap in Discover, Install autonomous agents, opens the agent catalog for you. Test Health shows “Autonomous agents this week” with a link, Go to agents.

Why use agents#

  • No red run stays unexplained. Daily triage groups yesterday’s failures by cause before the team starts work.
  • Everyone gets the same picture. Daily QA report and Weekly management report go to e-mail (with a PDF), Slack, Teams, Discord or a webhook.
  • You see what is not tested. Coverage gap report lists untested points on the app map and sub-requirements without test cases.
  • Deploys are checked right away. Deploy agents wake up when your CI reports a new deploy.
  • You stay in control. An agent proposes; a person applies. Every run, step, token and cent is logged.

How a run works#

  1. Something starts it

    You click Run now, the schedule comes due, or a new deploy arrives. A paused agent never starts. If the agent already has a run going, a second one is not opened.

  2. The budget is checked

    If the agent has already spent its daily budget today, the run is recorded as Budget blocked and does not start. See how spend is counted.

  3. The model works step by step

    Each step, the model decides what to do next and calls one tool. It may call only the tools ticked in the agent’s settings; anything else is refused and logged as DENIED.

  4. Changes become proposals

    When a tool would change something (a locator, a defect, a tag), the server does not run it. It records a PROPOSAL step with the exact arguments.

  5. The run ends

    The agent writes its summary and report. The run ends Succeeded, Failed, or Stopped at limit when it hits the step or cost limit. The report goes to the targets in Reports.

    You open the run, read the report and apply the proposals you agree with.

The building blocks#

Agent
One job with its instructions, allowed tools, trigger and limits. It belongs to one project. Every save creates a new version.
Autonomous agent
An agent installed from the catalog, with instructions written by SmartestQA. You can change its settings.
Custom agent
An agent you build yourself: you write the instructions and pick the tools. It starts paused and runs on demand.
Trigger
When the agent runs: Manual (only with Run now), Schedule (a cron expression) or After each deploy.
Run
One execution of an agent, with its status, steps, tokens, cost, summary, report and proposals.
Step
One entry in a run: MODEL (a decision), TOOL (a tool call), PROPOSAL (a change waiting for a person), FINAL or ERROR.
Proposal
A change the agent wants, saved with its exact arguments. Apply runs it as you; Dismiss drops it.
Runs on
Who executes the run: the SmartestQA server (uses the project’s AI connection and costs tokens) or your Claude Desktop (a scheduled task there calls the agent).

The screen#

Autonomous Agents page with the four KPI cards, the Agents tab and three agent cards (one running)
Autonomous Agents with three installed agents.

The tab is kept in the address (?tab=runs), so a refresh or a shared link opens the same tab. While any run is queued or running, the page refreshes itself every 5 seconds.

Before you start#

You wantYou needCheck it
Report agents (triage, QA report, coverage gap)Test runs and pipelines in the project; the project’s AI connection workingReport shows recent runs
Deploy verificationA code repository in Deploys & Repos and a pipeline tagged smokeThe catalog card shows no warning
Deploy impact analysisA code repository with an access token that SmartestQA can readThe catalog card shows no warning
Reports by Slack, Teams, Discord or WhatsAppThe channel set up and enabled in IntegrationsThe channel is tickable in Reports
Runs in Claude DesktopClaude Desktop with the SmartestQA MCP installed, open at the scheduled timeThe card shows no “Expected run did not arrive” line

Best practices#

Do

Don’t

  • Click Apply all without reading what each proposal changes.
  • Raise the daily budget to make a failing agent “try harder”: fix its instructions instead.
  • Tick every tool in a custom agent. Give it only what its job needs.
  • Install two agents that do the same job on the same schedule.

Agents, MCP or doing it yourself#

WayBest whenWho decides
Autonomous agentThe same check must happen every day or after every deployThe agent proposes, you apply later
AI assistant through MCPYou want to build or fix tests in a conversation, nowYou approve each diff in the chat
Report and Maintenance screensYou want to look at one run or one heal yourselfYou

In this section#

Didn’t find your answer?Write to support with your project name and what you see on screen.

info@smartestqa.ai
© 2026 Selenity Technology · SmartestQAsmartestqa.aiThis page as MarkdownChecked against smarttest-fe predev 2a8c79cc, smarttest-be 633c2a03 (aiagent package), smartest-mcp-server 1.31.1