Build your first agent in Perplexity Computer

This is the module where you build a real agent. We're doing it in Perplexity Computer — the environment you're already in — because it hands you tools, skills, connectors, subagents, and cron out of the box.

The four primitives

Computer has four building blocks. Master these and you can compose almost anything.

Build: a weekly lead-research agent

Goal: every Monday morning, find new sales opportunities relevant to your business. Score them for fit. Post a Slack summary. Save a doc with details. Same architecture works whether you sell government contracts, real estate, consulting, or artisan cheese — you just change the sources.

Step 1 — Spec it in plain English

Before typing a prompt, write the brief. Every good agent starts here.

Brief · lead-radar.md# Lead Radar

Goal: surface new sales-qualified leads and opportunities so I can
decide what to pursue by 10am Monday.

Sources (customize to your business):
  1. Industry-specific bid boards or lead sites
  2. Google Alerts for keywords that match your services
  3. LinkedIn for named target accounts (if you're B2B)
  4. Local news / permits / filings (if relevant)

Scope filter: match the services you offer. Ignore anything outside
your focus.

Budget filter: minimum project size that makes sense for you.

For each match, extract:
  - Source, title, due date, budget (if stated)
  - Scope summary (3 sentences)
  - Fit score 1-5 (5 = perfect for us)
  - Link to the original

Output:
  - Google Doc "Lead Radar - YYYY-MM-DD" in "/Business Dev"
  - Slack post to #new-business with the top 3 by fit score
  - Never auto-reply or auto-submit anything

Step 2 — Turn it into a scheduled task

Open a fresh Computer thread and say:

Computer will:

1. Read the brief from Drive (uses the Drive connector).
2. Spawn browser subagents to hit each source (parallelization pattern).
3. Extract structured JSON per RFP.
4. Score for fit using a scoring prompt.
5. Create the Google Doc (Docs connector).
6. Post to Slack (Slack connector).
7. Schedule itself to repeat weekly.

Step 3 — Promote the routine to a skill

If you're going to run this every week, package it as a Skill so it's reusable and versioned. In Computer, ask:

"Turn the Lead Radar workflow into a saved skill named lead-radar in my user scope. Include the brief, the scoring rubric, and the exact output format as reference files."

Now any future thread can say "load the lead-radar skill and run it on today's date" and get the same behavior.

When to spawn subagents

Subagents let you fan out work in parallel. Use them when:

• You have 5+ independent research targets ("visit these 10 sites and summarize each").
• The main context is getting cluttered with raw material you don't need long-term.
• Different steps need different expertise (research vs. writing vs. design).

Pattern: the main agent plans → spawns subagents in parallel → each writes findings to a file → main agent reads the files and synthesizes. This is exactly the "orchestrator-workers" pattern from Module 4.

Anti-patterns to avoid in Computer

• Over-scoping the first task. Start with one source, one output. Iterate.
• Forgetting the human checkpoint. For anything that emails, publishes, or spends money, require confirmation.
• Not using skills. If you'll run it more than twice, save it. Skills are how you compound.
• Running everything hourly. Cadence matters — every hour is expensive and noisy. Daily or weekly is right for most things.