Agentlas
Agentlas
Search docs…⌘K
← Back to app
IntroductionUse AgentlasAgentlas OSGuidesTrust & safetyOperationsLegalChangelog
Guides

Build an agent or team

Turn a repeatable job into an agent package you can inspect, install, and reuse. Start with the result you need, the inputs it will receive, and the tools it may use.

What building an agent means

A build creates reusable instructions and supporting files for an agent or team. It does not complete every future job for that agent. Test the package on a small real example before relying on it.

Single agent
One specialist with its own instructions and references.
Multi-agent team
Several roles with a coordinator and a plan for handing work between them.
Package existing agent
Convert or repair an existing local agent into an Agentlas package.

1. Build from a runtime — Claude Code, Codex, and others

Install the Agentlas OS integration for your host, then check that its builder command or skill is available. Command spelling depends on the host; use the skill picker in Codex and the slash-command picker in Claude Code.

Install Agentlas OS for your host →
Claude Code:
/hep-build Create a research analyst that compares supplied sources and writes a cited report

Codex skill:
$hephaestus-build Create a research analyst that compares supplied sources and writes a cited report

Agentlas Terminal:
agentlas build --help
agentlas build "Create a research analyst that compares supplied sources and writes a cited report"

Answer the interview with examples of useful output, expected inputs, and actions the agent may take. Attach your references where the host supports it. A larger team or more research can take longer; follow the build's progress and questions instead of expecting a fixed completion time.

2. Build from Agentlas Desktop

  1. In Desktop's left menu, open Agent Forge → Build.
  2. Choose Single agent, Multi-agent team, or Package existing agent. Describe the job, completion criteria, inputs, and allowed actions. For packaging, attach the existing folder.
  3. Choose an output folder and a connected Build model, or use Auto (active engine). Add example files or a reference folder if needed.
  4. Select Start build (deep interview), then answer the requirements questions.
  5. Review proposed MCP connections. Attach the connections you want, skip an unavailable connection, or continue without MCP. Supply required credentials through the app's connection settings.
  6. Follow Confirm brief, Engine & MCP, Research & plan, Generate package, Verify, and Deliver. If the app offers a higher model, review its usage implications before choosing it.
  7. Open the generated files and findings. Use Open in My Agents or Open org chart to inspect the installed agent or team.

You can stop a build. When a recoverable failure offers Resume from saved files, use it to continue from the retained folder. Choose New build when you intend to begin another build.

3. Save or publish the package

After local installation, choose Upload to Cloud for private storage or Keep local only. Public Hub distribution is a separate action. Review the source-access choice and the public package before submitting it.

Agent Cloud stores a package; it does not host its model or run it for you. Restore and install it on another Desktop, then run it with that Desktop's connected model and tools. A paired phone reaches the agent through that Desktop.

Set the package's intended name in the interview. Desktop also lets you change an installed agent's local display name; this does not rename the published package identity.

Test your agent on a first task →
Save, publish, and restore agents →
Build an agent or team · Agentlas docs