Content settings
Attach an RSS feed so the agent writes posts based on new articles, or let it generate topics autonomously.
Tip
Giving your agent specific examples of your best posts dramatically improves output quality.
What are AI agents?
AI agents are autonomous programs that generate and publish social media content on your behalf based on instructions you define.
Agents can pull from RSS feeds, trending topics, your product updates, or any custom data source you connect.
Creating your first agent
Setting up an agent takes about five minutes.
1
Go to AI Agents
Open the AI Agents section from the sidebar and click New Agent.
2
Define your agent
Give your agent a name, choose profiles to post to, and set the posting frequency.
3
Write the instructions
Describe your brand voice, topics to cover, what to avoid, and formatting preferences.
Workflows
Workflows let you chain actions together using a visual node editor.
Each node is an action: Generate content, Post to profile, Wait, or a conditional branch.
Safe mode
With safe mode enabled, all agent-generated posts go into a review queue before publishing.
Important
Enable safe mode when first deploying an agent. Review a few batches before switching to fully autonomous publishing.
AI agent failure modes
Agents fail in predictable ways. Here are the four that bite new users — and how to avoid each.
- Vague brand-voice instructions. 'Be friendly' produces generic content. 'Use first-person voice, mention specific tools, end with a question, max 200 words' produces consistent content. Treat instruction-writing as serious investment — your agent's output quality scales with instruction precision.
- Disabling Safe Mode too early. Auto-mode looks faster, but agents make their dumbest mistakes in the first 50 outputs (mismatched tone, hallucinated facts, off-brand examples). Stay in Safe Mode for the first 2 weeks; auto-mode after you've validated quality.
- Token budget set too low. Cheap drafts (Claude Haiku, GPT-4o-mini) work fine for short-form social posts but produce shallow long-form content. If your LinkedIn drafts read thin, raise the per-agent token budget and switch to a more capable model.
- Treating an agent like a setup-once-forget-forever tool. Agents drift as platforms evolve. Schedule a monthly review: read 10 recent outputs, update business context if your positioning changed, and tweak instructions based on what's underperforming.
Three agent patterns that work
Battle-tested deployments — copy these structures rather than designing from scratch.
Pattern 1: Daily content drafter (Content Planner)
Cron: every weekday at 8am. Goal: produce 3 platform-tuned drafts (Twitter, LinkedIn, Bluesky) from the latest RSS-feed article. Token budget: 2000 per run. Approval: required (Safe Mode on). Output lands in your inbox before your morning coffee; 5-minute approval loop.
Pattern 2: Hourly lead scanner (Lead Hunter)
Cron: every hour, 9am-6pm weekdays. Goal: find up to 5 high-intent leads matching keywords. Score floor: 70. Token budget: 1500 per run. Reply drafting: enabled (drafts queue, you approve). Daily output: 30-50 fresh leads ranked by intent, ready for your review.
Pattern 3: Weekly performance digest (Analytics Agent)
Cron: Sunday 10am. Goal: pull last 7 days of metrics across all connected platforms, identify top performer + biggest miss, write 1-paragraph summary. Token budget: 3000 (it's a longer thinking task). Output: email digest. Replaces the Monday-morning dashboard scroll.
AI Agents FAQ
Which AI provider should I use?
Mix and match per-agent. Cheap drafts → Claude Haiku ($0.001/1K tokens, fast). High-stakes content → Claude Sonnet or GPT-4o ($0.003-0.015/1K, slower but better tone). Anything below 100 tokens → consider GPT-4o-mini for cost. Mirgent supports per-agent provider selection so you don't have to compromise.
How much does running an agent actually cost?
Typical solo creator with all four agents on default settings: $5-15/month in AI costs (BYOK, billed by your provider). Agency running parallel agents for 10 clients: $50-150/month. Costs scale linearly with agent runs × token budget — set hard caps per agent so worst case is one bad day, not a runaway $500 surprise.
Can agents post directly to my accounts without approval?
By default, no — every public action goes through your approval queue. You can opt individual agents into auto-mode per goal (e.g. 'auto-post Content Planner drafts that score 90+ on internal quality check'). Auto-mode is logged separately; you can audit any post the agent ever made.
What happens to memory when I edit an agent's instructions?
Memory is preserved (the agent still remembers what it did and how those outputs performed). Instructions change how it generates — same memory, different rules. This is intentional: you can tune the strategy without losing learned context. If you want a clean slate, clone the agent and start fresh.
Can I build a custom agent for a niche use case?
Yes (Pro and above). Custom agents are defined by goal + tools + cron — same primitives as built-ins. Examples we've built: hashtag trend monitor, competitor activity tracker, automated brand-mention responder. Custom agents share the same memory + business-context + approval-queue infrastructure as built-ins.