GPT-5 agents: workflows that run themselves.
GPT-5's agent capabilities let ChatGPT run multi-step workflows autonomously — visit websites, fill forms, take actions. It's the most powerful agentic AI in consumer software today, and the most risky if misused. This lesson covers what agents do well, the supervision patterns that prevent disaster, and the cost-control playbook.
The mental model
Agents are interns with internet access. Treat them accordingly.
GPT-5 agent mode can do remarkable multi-step work autonomously. It can also confidently complete the wrong task, spend tokens fast, and make irreversible changes. The skill isn't using agents — it's supervising them well.
Never give an agent access to anything you wouldn't give to a brand-new intern. No banking. No production systems. No accounts where one wrong action matters. Use throwaway browsers, throwaway accounts, public data. The capability is real; the trust required is still being earned.
Workflow 01 Research and synthesis at scale
The 30-source research session
Give the agent a research question. It visits 20-30 sources, synthesizes, and reports back.
The prompt that works
Best use cases
- Pre-meeting prep with depth
- Industry analyses
- Competitive intelligence
- Trend reports
Workflow 02 Multi-step web tasks
Find, compare, recommend
Tasks that require visiting many pages, comparing options, and recommending one.
The prompt that works
Best use cases
- Vendor comparison and selection
- Product research with specific criteria
- Travel planning with constraints
- Anywhere you'd otherwise compare 10+ options manually
Workflow 03 Periodic monitoring
Set-and-monitor weekly tasks
Schedule an agent to perform a weekly task: check 5 specific URLs for changes, summarize industry news, audit accounts.
The prompt that works
Best use cases
- Competitor monitoring
- Industry news roundups
- Inventory or pricing tracking
- Periodic compliance/audit checks
Final challenge: one supervised agent task
Identify one research-or-monitoring task you do manually that takes 30+ min. Use GPT-5 agents for it. Watch it carefully the first time. Verify every claim. Iterate on the prompt. If it works, schedule it. If it doesn't, learn why and try a different task type.
What you can do now
- Understand what agents do well vs. poorly
- Apply safety patterns (throwaway accounts, supervised first runs, no irreversible actions)
- Use agents for research, comparison, and periodic monitoring
- Set cost budgets and verify before acting on agent output
- Recognize when scripting (Python/zapier) would be more reliable than agents