Building AI agents

What “agent” means here

An agent is a named AI assistant with a purpose, a set of allowed tools (search, fetch client data, etc.), and optional scheduled runs. Think of it as a specialist on your team you can talk to or assign a recurring job.

Prebuilt agents

• News agent — search headlines, summarize breaking stories, find articles relevant to your clients.
• Trend agent— analyze what's moving in your trend feed, explain why, suggest pitches.
• Pitch agent— conversational pitch drafting. Good for iterating on a draft (“make it shorter”, “add a stat”).
• Report agent — generate on-demand reports across a date range or client.
• Outlet Researcher agent— find reporters by beat, summarize what they've covered recently.
• General agent — open-ended chat with access to your full workspace.

When to build a custom agent

Build one when you have a repeating workflow that the prebuilts don't do. Examples from our customers:

• Competitor Watch — daily scan of 4 competitors with a one-line summary of what changed.
• Executive Brief — weekly digest written for C- suite, pared down from raw mentions.
• Crisis Signal — flags sentiment drops + PR emergencies in Slack before the comms team notices.

Pro and Enterprise include custom agents. Starter does not.

How to build one

1. Go to Agents → + New agent.
2. Name + purpose — what the agent does in one sentence.
3. Instructions — how the agent should behave. This is the system prompt. Be specific.
4. Tools — tick what this agent is allowed to do. See below.
5. Schedule (optional) — if it should run daily / weekly, set the cadence + where to deliver results (email, Slack, or just in-app).

Tool permissions — what each does

• Web search — live Google / Reddit / news search for current information. Costs more per call; toggle off for cost-sensitive agents.
• Read clients — access to your client list, profiles, and brand voices.
• Read mentions + trends — the live feed.
• Read pitches — draft history.
• Write pitches — create new drafts directly.
• Send Slack — post results to a configured channel.

Tighter tool permissions = cheaper per-turn + lower risk. Start with the minimum and add as the agent falls short.

Running an agent

• Chat— open the agent's page and type. Streaming responses with tool-call status chips.
• Scheduled — runs automatically on its cadence. Results land in-app, email, or Slack based on your config.
• Triggered(Enterprise) — fire an agent from a webhook. E.g., “new mention with negative sentiment → run Crisis Signal agent.”

Cost control

Agent calls count against your monthly AI credit budget. Especially chatty agents (long back-and-forth, web search) burn credits faster. Three things that help:

• Disable web search on agents that don't need it
• Keep instructions tight — verbose system prompts cost tokens on every turn
• Enterprise: bring your own Anthropic API key and PRplus doesn't touch your plan budget at all