Configure a Seren Employee

Choose the deployment path

Managed prompt or custom bundle

Use deploy-prompt for a Seren Employee backed by a managed agent with presets, approvals, revisions, and cloud execution. Use the bundle deployment path when you need your own runtime, entrypoint, dependencies, or packaged code. Both paths run on Seren Cloud, but only the managed path exposes the resolved employee specification through Seren Agent.

Step 1

Deploy a read-oriented employee

Start with the narrowest useful policy. The research_monitor template is suited to research and monitoring, while workflow_agent supports action-oriented work. Tool presets select groups of capabilities. The approval policy controls whether mutating publisher and MCP calls are available.

seren agent deploy-prompt --name "Earnings Monitor" --template research_monitor --tool-preset live_data,database --approval-policy read_only --model-policy balanced --prompt "Monitor quarterly filings for companies in the research queue. Summarize changes in revenue, margins, guidance, and material risks, and cite the source documents."

The model policy chooses among fast, balanced, and deep profiles. To pin a specific model instead of the platform default, list the available models with seren agent private-models list and pass its identifier with --model-id.

Step 2

Inspect the resolved specification

Inspect the managed deployment and its immutable revision history before changing it. The resolved view includes the selected presets, effective approval policy, visible resources, remote delegation allowlist, and any attached eval gate.

seren agent managed-get <deployment-id>
seren agent managed-revisions <deployment-id>

Working data

Let skills declare database storage

A stateful skill can declare the database it needs and choose organization or user scope. Seren creates or reuses the database when the employee is deployed and keeps it across restarts and updates. You do not need to create a database project or add a connection string to the employee. The platform manages the database lifecycle and credentials, while SerenDB encrypts its underlying storage at rest.

Choose organization scope if several employees or a swarm need to coordinate on records that organization members can use. Choose user scope if each user needs separate working data for the same skill. A skill's stable slug identifies its database. A change to the displayed title does not move the data.

Browser work

Run Playwright skills in Seren Cloud

A browser-enabled employee can run existing Python and Node Playwright scripts and Playwright MCP workflows without converting them to Selenium. Seren attaches the script to a fresh managed Chromium session, keeps ordinary network access out of the skill process, and limits browser navigation to the destinations declared by the employee policy.

Authenticated browser state is encrypted and reused for the same employee and skill. Downloads remain temporary unless the skill explicitly moves them into approved storage. This keeps login state available across runs without turning the browser profile or download directory into durable general-purpose storage.

{
  "approval_policy": "allow_mutations",
  "capability_policy": {
    "skills": {
      "enabled": true,
      "execute_scripts": true
    },
    "browser": {
      "enabled": true,
      "profile": "full"
    }
  },
  "runtime_policy": {
    "version": 1,
    "network": {
      "default": "deny",
      "egress_rules": [
        {
          "host": "example.com",
          "port": 443,
          "protocol": "tcp",
          "enforcement": "enforce"
        },
        {
          "host": "*.example.com",
          "port": 443,
          "protocol": "tcp",
          "enforcement": "enforce"
        }
      ]
    }
  }
}

Save the policy in a JSON file and pass it with --agent-config employee-browser.json when deploying or updating the employee. The full browser profile is required for Playwright scripts. Script execution also requires an attached skill bundle and the allow_mutations approval policy.

Add every website, identity-provider, redirect, API, and asset hostname the workflow needs. A wildcard such as *.example.com does not include example.com, so declare both when the workflow uses both. Seren supplies the browser endpoint and its internal credentials. Do not add them to employee configuration or skill secrets.

Step 3

Preview policy changes

Preview an update before applying it. Remote A2A delegation remains disabled unless the employee has an explicit origin allowlist. Entries can be complete origins or hostnames. If the employee must not call remote agents, leave the list empty.

seren agent managed-preview <deployment-id> --allow-remote-agent-origin https://agents.example.com --approval-policy allow_mutations

seren agent managed-update <deployment-id> --allow-remote-agent-origin https://agents.example.com --approval-policy allow_mutations

Step 4

Require a fresh passing evaluation

An eval gate blocks scheduled and API-triggered runs unless the selected eval set has a recent passing verdict. Set the allowed result age in seconds, preview the change, then apply it. Clear the gate when it is no longer part of the employee's operating policy.

seren agent managed-preview <deployment-id> --eval-gate-set-id <eval-set-id> --eval-gate-max-age-seconds 86400

seren agent managed-update <deployment-id> --eval-gate-set-id <eval-set-id> --eval-gate-max-age-seconds 86400

seren agent managed-update <deployment-id> --clear-eval-gate

Advanced tools

Set per-tool limits when needed

Managed deployments accept --agent-config with a path to a JSON file for advanced tuning of the managed fields. Raw tool_definitions belong to custom bundle deployments: pass --orchestration-config <path> to seren agent cloud deploy. A tool definition can override its execution timeout and cap captured output without changing the global employee policy.

{
  "tool_definitions": [
    {
      "name": "fetch_large_report",
      "description": "Fetch and summarize a large report",
      "timeout_override_seconds": 120,
      "max_output_bytes": 16384
    }
  ]
}