Skip to main content

Save your first agent

An agent is a saved instruction with a goal, autonomy level, and optional triggers. Unlike a one-off chat question, agents are durable: every run creates a persisted task with a full plan, tool trace, and result.

Before you start

You need a SixDegree environment with at least one integration connected and discovery completed. If you haven't done that yet, see your first environment.

Choose a good first goal

Pick something small, repeatable, and read-only. A solid first agent:

Goal: "Summarize all accounts renewing in the next 30 days that have an open support escalation, and list the owning rep for each."

This is observe-only: no writes, no destructive actions, no approvals. It's the kind of question you'd normally ask in chat, but saved so you can run it every morning without typing it again.

Start observe-only

Your first agent should be read-only. Once you trust what the agent does, you can raise its autonomy tier to let it take action.

Create the agent

In the dashboard, navigate to Agents and click New Agent. Fill in these fields:

FieldWhat to enter
Namedaily-renewal-risk
DescriptionSummarize accounts renewing in 30d with an open escalation
GoalThe instruction above
Autonomyobserve: reads only, no writes
ModelLeave blank to use the environment default

Leave Scope and Blast Radius at their defaults. The defaults give the agent read access to everything in the environment and a sensible ceiling on cost and tool calls.

Click Save.

Run the agent

Click Run on the saved agent. The dashboard opens a live task view with four panels:

  • Plan: steps the agent declared, checked off as they complete.
  • Tool calls: every tool invoked, with arguments and results.
  • Usage: input/output tokens, total cost, and wall-clock time.
  • Result: the final summary the agent produced.

Each run is a task: a durable record of what the agent did. The plan, trace, and result are persisted so you can review any run later.

Tasks are permanent

Completed tasks are never deleted. You can return to any run and see the exact tool calls and results the agent produced.

Add a schedule

To run the agent every weekday morning instead of manually, add a scheduled trigger.

From the agent detail page, click Add Trigger and select Scheduled:

FieldValue
Namedaily-renewal-risk-morning
Cron0 9 * * MON-FRI
Run asYourself: the agent executes with your permissions
0 9 * * MON-FRI

This fires the agent at 9am on Monday through Friday. Each fire creates a separate task with its own full run history.

Understand autonomy tiers

This first agent uses observe: it can read but not write. When you're ready to let an agent take action, raise the autonomy tier. The platform routes writes accordingly:

TierWhat the agent doesApproval required
observeRead-only. No writes attempted.None
suggestProposes writes; nothing executes until approved.Every write
actWrites execute automatically. Destructive actions require approval with MFA step-up.Destructive actions only
act_destructiveFully trusted. No gates.None
Raise autonomy deliberately

act_destructive removes all approval gates. Only use it for agents you have observed running correctly over many runs.

What to build next

Once your first agent is running, try these patterns:

  • Triage agent: "When a Zendesk ticket is opened with the urgent tag, trace it to any related incident and the account's revenue, then post a summary." Fire from an event trigger.
  • Webhook responder: "When Stripe posts a payment failure, look up the account entity and gather its open opportunities and recent activity." Fire from a webhook trigger.
  • Incremental scanner: An agent that uses memory to record where it left off so each run picks up where the last one stopped.

Next steps

  • Agents: autonomy tiers, scope, and blast radius explained
  • Triggers: event, scheduled, and webhook firing
  • Approvals: how writes get routed to a human
  • Built-in tools: plan, memory_*, and the rest of the built-in toolkit
  • Tasks: understanding run history and the task lifecycle