AI tools are constantly changing and are getting more powerful with each release. Since creating my first custom agent The Super-Investigator , it has been incredibly tempting to use custom agents for every situation, but in reality, not every problem needs one.
Knowing when to build a custom agent has been the most challenging and I feel like I'm still maybe under utilising the option so that I don't over complicate my work.
The Custom Agent Decision Checklist:
- Repetition: Is this a prompt that I am repeating daily or weekly to perform a similar task?
-
Complexity:
Am I switching between multiple modes like
planandagentconstantly? - Consistency: Do I need the output to follow a particular structure each time I use this prompt? Is there a set of tools that I need to use to achieve this?
- Context: Is there a specific set of context that the agent needs every time I utilise this prompt?
Why I chose to write my first agent:
I wrote my first investigator agent a few weeks ago to solve the problem of repeated prompts. I was constantly asking Copilot to complete an investigation for me which, at the start, involved the same prompt and the same context no matter what I needed investigating. I was copying and pasting the majority of my prompt and updating only the small parts that added context to that day's particular problem.
Once I started adding in MCP tools like Atlassian linking context to Jira tickets and saving investigations in Confluence, the task became quite tedious as I was switching between plan for the initial investigation and agent for any updates and to save the plan to Confluence.
This was an incredibly tedious process and although I was saving a lot of time having Copilot do the initial investigation, I was still wasting a lot of time switching between built-in tools. Having the custom agent meant that all the shared detail in the prompt could move to a custom agent file that could be used multiple times by anyone in the team. I could also add in automation to link to Jira and Confluence without having to add an additional prompt or wait until the investigation was complete.
Working with a custom agent means that my team is now generating structured documentation by repeating the same investigation pattern. We have the capacity to have a unified structure for investigation documentation and links to Jira tickets.
When I utilise the built in Plan/Agent workflows:
There are other instances where I have used a prompt frequently and have created agents (story for another time) but the majority of the time I use the built in Copilot functionality of plan and agent.
My investigation agent has become step one in most of my coding sessions and I use that as a base to get Copilot to plan what to do next. Because the investigation comes with code snippets and detail on what is likely to change, I can use plan easily as the next step, iterate on that and then assign to an agent to go ahead to build and review the code.
While I could add a loop in my Investigation agent to automatically move to the plan and build stages, I don't believe that in my scenario it's needed as I still like to make changes and update the plan as I go. While this is maybe a next step I could look at, I would need my plans to be more accurate and require less human intervention before it's something I look into.
Built-in tools are there for a reason, they're common used functionality for everyone at any level.
Final Thoughts:
So I leave you with this; custom agents are incredibly powerful, but only when applied to the right situations. They still require time to build and to ensure they are working as expected and if they're going to sit in a folder and be used once every few months, I'm not sure it's worth it.
For me, I find that as soon as I need to look back through my prompts to copy and paste something into the context window, it's time to go through my checklist to see if I can utilise a custom agent.
There is definitely a case-by-case argument to be made for custom agents but as the tooling and the models improve, I'm not sure if the future is custom agents or more skills focused to extend the built-in tooling.
The question isn’t what AI can do—it’s when it’s worth structuring how it does it.




