Good Agents
Like bad prompts, bad agents can ruin an automation, causing it to operate inconsistently or erratically.
What makes a good agent?A good agent has instructions that are applicable to multiple prompts within the same domain, like "typescript programming", "campaign planning", "database administration", and "marketing image asset creation".
A good agent has instructions that are NOT going to contradict the instructions in the prompts the agent evaluates.
A good agent uses the most suitable model, like gpt-4.1 or o4-mini, to
correctly carry out the instructions of an automation.
A good agent, like a good prompt, avoids instructions that in reality don't help, such as:
- Don't make mistakes!
- Don't hallucinate!
- Do what I mean!
- I'll give you $500 if you get this right!
A good agent does NOT have specific task instructions unless it makes sense to carry out those task instructions with every prompt the agent evaluates.
For example, imagine this specific task instruction in the middle of the agent's overall instructions:
...
Use the tool `postgres_exec` to query the database for system settings, using
the query: `SELECT ...`
...When adding task specific instructions into an agent's overall instructions, ask yourself this question:
Is this something reasonable for the agent to do in all the prompts I will have this agent evaluate?