Look, y’all can make fun of me all you want. I know the industry has been talking about “agent this” and “integrate agents with that” for what feels like EONS at this point.
(It’s been maybe more than a year lol)
But every single time I’ve sat down to try to understand how powerful agents are, how to connect them to services through MCP, what context windows are, how to discover and install skills, how to use them in my own workflows, etc...
Basically, I had access to all these amazing automation tools at my fingertips and understood how they worked... but had no idea when, where, or why an agent would be useful in my life.
For example: I knew that I could connect my AI agent to Gmail, so it could read and write emails for me. That’s really cool and all... but what would I use that for exactly? To summarize my emails? To eliminate spam? To write to certain people automatically according to some parameters in a prompt? I didn’t know or have a use to connect to Gmail in that way.
Now imagine learning about all these cool new tools and integrations available with agents... and not knowing what to use them all for.
This is such a specific problem that I’m positive a lot of devs are going through, but can’t exactly articulate.
What I was missing was a “mental model” for using these tools in a context of improving workflows and creating automations.
So gather round! This hopefully will turn into an email JAM-PACKED with insights around agents for those who still don’t “get it”, the same way I didn’t just a few weeks ago.
1) If you want to start using agents in your life/work, start by detailing your workflows.
2) Once you understand your workflow, think about how you could improve/optimize parts of it.
3) Only after doing those two things, THEN you think about agents/integration/mcp/skills/context/implementation
Yes, I know — I KNOW how incredibly simple this sounds, but this is NOT obvious. The devil is in the details here, and waaay better explained through an example. Let’s take a standard developer workflow and go step by step to see how we could use agents.
Let’s say you’re a Frontend developer who works with Jira tickets, GitHub, and Figma. Cool. What would your workflow look like?
It might look something like:
1) Grab the ticket you have in Jira assigned to you with the requirement and Figma design.
2) Look at the Figma design and develop it in the code.
3) Push code to Git and create the Pull Request in GitHub.
And... that’s it, right? That’s the typical list of steps you’d imagine developers experien-
WRONG. We didn’t do a good job detailing the workflow.
As I said, the devil is in the details. If you want to really start finding opportunities to optimize and automate parts of your work, you have to go two levels deeper.
THIS is the level of detail I’m talking about:
1) Open the browser and go to the Jira website.
2) Navigate to the sprint dashboard and select your ticket
3) Look at the ticket and follow the link to the Figma design
4) On another screen (let’s say you have multiple monitors), open up your IDE and check out the main branch.
5) Create a new branch with the Jira ticket’s number and start working, committing your changes with whichever commit standard agreed upon by your team.
6) Once you’ve finished, push your changes to your branch.
7) Navigate to GitHub, find your repository, and click on “Pull Requests”.
8) You create the PR with a title, description, assigned reviewers, etc.
You get the picture by now. This is what I mean by “detailing your workflow”.
I realized that to apply agents (or automate anything in general) into something useful you’d use in your work, you have to be able to express how you get your work done in the first place. This forces you to take a step back and observe some of the habits and processes you’ve developed from working over the years. It’s only through this process that you realize that, yes — there are small annoyances or waiting times you endure throughout it. And those are the perfect places to look through automation!
Once and only when you’ve done this detailing of your workflow can you move on to...
Now that we have our workflow detailed in writing, let’s find out how we could improve it!
(I’m not going to be fancy here, and only illustrate this point with small optimizations)
What I notice is that steps 1-3 are about navigating to the sprint dashboard in Jira. Hmm, how could we make that faster or more optimized? Could we add a quick browser bookmark link to our sprint dashboard so we’re one click away? That could save some time.
Steps 4-6 are about getting our programming done. We probably make use of AI a lot here, so let’s imagine we’ve experimented plenty in this area.
Steps 7-8 are again about navigating webpages and setting all the details related to our pull request. Hmm, maybe we can get the AI we use to read our code to help us generate a clean PR description to copy-paste and save some time? Perhaps we could also set another browser bookmark to our repo’s pull request page as well?
The idea here is to get those brain juices flowing. What we’re doing here is analyzing which aspects of your workflow can be sped up or optimized. We’re questioning what the best way to do something is in each step. This also helps us put some distance between us and our workflow. This makes the last step a little easier when we...
In this step, forget only optimizing and speeding up work — let’s remove or even replace annoying parts of our workflow altogether. Here is where some basic AI / Agentic knowledge comes in.
Agents can be described as AI that can execute tasks and run commands.
Skills are a collection of tasks, commands, and context you can bundle into one single command that you can make an agent execute.
You can integrate these agents with other tools and services, a lot of times with built-in connectors or something called an “MCP server”.
Let’s avoid the technical talk for now and keep it high-level. (You don’t need to know what MCP is right now, for example. Just that you can integrate agents with other tools. You can google/chatGPT the rest later 😜)
This lets us ask some interesting questions about how best to use these tools. Questions like:
Which parts of my workflow can I streamline, replace, or automate away?
Are there any friction points in my work that don’t require critical thinking?
What services or tools do I work with currently? Are there integrations with those tools or services? What could I hypothetically do with those integrations?
Can I reduce the amount of typing, clicking, or navigating I currently do to get my work done?
Which are the parts of my workflow I absolutely cannot automate away?
What are the tradeoffs or risks of introducing AI into parts of my workflow? How do I reduce them?
We’re trying to use our noggin’ here and experiment with the tools at our disposal to see what’s possible. But more importantly, when and when shouldn’t we use agents for what parts?
Steps 1-3 were about navigating to the sprint dashboard in Jira. Could we avoid going to the Jira website altogether? Why yes, yes we can! We can connect Claude directly to Jira so that it brings the ticket information whenever we ask it. That saves us quite a bit of time!
(I didn’t know about this 3 weeks ago, and it was one of the things that made this agent stuff “click” for me)
We can also connect our agent to Figma through another integration, so it can “see” the design we need to implement. Which means our agent now has both the context from the ticket + the design.
Steps 4-6 were getting our programming done. Since the agent we’re working with already has the context + the Figma design, we save time having to type up what needs to be done to the AI.
Steps 7-8 were all about creating the pull request. Could we integrate it with GitHub? You bet we can! Which means that again, we don’t have to open a tab and painstakingly click through a bunch of buttons or type a lot to create a PR. We can detail to the agent how we want the PR to look, and it can handle it for us with some supervision.
What’s cool about agents is that they can string multiple commands, steps, and tools together. So you could, for example, create a skill that:
1) Grabs the ticket assigned to you in the current sprint
2) Reads the ticket and grabs all the context surrounding it, alongside the design
3) Creates an implementation plan which, when accepted, starts to put itself to work
4) Automatically creates the branch, updates the code, and tests it against the current project and team standards
5) Commit and push the code to the branch under certain commit message standards
6) Automatically creates the PR for you once it is done.
Which is wild to me still. And perfectly explains a lot of the agentic AI hype, as well as the whole “engineers will be out of a job” theories we’ve been hearing so much around.
(Topic for another day)
Either way, I hope this email helped at least a few of you. It seriously was a problem I’ve been tackling to “get” for months until I worked with a co-worker who kindly answered all my basic questions until it all clicked.
(To be honest, I didn’t start this email with the thought “I’m going to write an article in email form today :D” but it just sort of happened. I realized that this was too important a topic to talk about to leave as just a small “3 tips I learned about agents” email.)
I just don’t think we talk enough about how to think through how or when to use these tools to create value. And since that’s one of the major themes I’m building this newsletter around, I felt it was my duty to share my insights with you all.
Let me know if you learned anything from this one! ✌️
P.S.
There are a lot of nuances, security concerns, privacy problems, and whatnot that I conveniently left out in this email around agents and AI, since I wanted to focus on what you can do with these tools. Just because I didn’t mention them doesn’t mean they don’t exist. Please double-check that everything you do is legal, secure, and keeps privacy in mind. I love y’all, but I am not responsible for your decisions 💖
Mauro Accorinti
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.