Yes I’m aware that I start sounding like an AI-pilled maniac, but I recently did a drastic overhaul of my entire workout and gym strategy with a team (yes, team) of AI agents as my personal coaches.
The gym is the last place I expected to incorporate agentic AI, but it has been going well enough that I now want to write about it!
Some stats of this experiment:
Turns out, AI is good with numbers
I’m currently neck deep into autonomously optimizing my entire portfolio with AI: Reading through product analytics, analyzing Google Search rank, optimizing keywords and ASO, to name a few.
So on a random day when I started my usual workout routine, I got curious: What would happen if I dumped an entire CSV export of all my recent workouts into Claude and let it tell me something interesting.
Well, besides giving me a few pretty graphs, Claude also tore through my workouts to tell me all the things I suck at.

This, uh, wasn’t great. I thought I was doing somewhat okay but here we are and I’m getting scolded by an AI.
Claude being Claude, it then made things sound like if I don’t change anything right now in this very moment, something terrible is going to happen.

Besides giving me life-ending panic, the output was useful: It identified clear gaps in my current training plan with direct suggestions how to fix them. All from uploading a single CSV export.
So I went on with my workouts, but added RDL, Face Pulls and Lat Pulldowns.
I continued repeating this “review” (nice way of describing Claude giving me panic) every now and then to see if I was doing better, and hey! I was!
But in the end, this was still my workout, with Claude recommendations sprinkled in.
My workout tracker has an API
I wanted to go one step further and give Claude full access to my entire workout data to see what would happen, so I hooked up Hevy, my workout tracker of choice, to a reverse proxy and started inspecting the app calls. I let Claude analyze the captured traces only for it to point something dumb out: Hevy has an API. Even better, there’s a community-built MCP Server around that API!
I quickly deployed the Hevy MCP to MCP Nest, a cloud service for hosting stdio/uvx MCP server and hooked it up to Claude.
Equipped with this new ability, I allowed Claude to plow through my Hevy data and after a bit of back and forth chatting, it told me my entire workout structure needed an overhaul.

I pressed it on more details and it gave me a clear example of a progression I should be using and that my current progression is trash.

So what I was doing is called a pyramid: For most lifts, I went up in weight until I hit the peak (eg 253kg on Leg Press), then slowly de-load and go down. Lots of sets and lots of volume, as seen by the 10 sets of Leg Press.
What Claude wants me to do instead is: 1x warmup, followed by 1x ramp, followed by 3-4 work sets on max weight. That’s it. No 10 sets, just 4 on max volume. It then went ahead and called my current progression “junk volume”.

That sounded so off to me, I thought Claude must be hallucinating some of it, so I did what I always do to reduce single-model hallucination: I asked another model for review.
I dumped the entire plan including previous data and imported it into Gemini 3.1 Pro, expecting it would disagree with the Claude plan but nope, it fully supported it:
The proposed restructuring correctly identifies and eliminates junk volume. Transitioning from a disorganized pyramid scheme to a strict double-progression model with designated work sets will immediately improve the stimulus-to-fatigue ratio.

It did adjust a few things though: 3 sets at max volume instead of 4, and made the last set a failure set (a set without cap on reps, you continue pushing until you can’t do a single more).
Armed with my new workout plan, I went in the next day fully expecting to breeze through those 3-4 work sets, after all I was used to 10-ish sets per workout, right? Well I quickly learned that both AIs working together were right and I wasn’t (oops!): The prescribed workout crushed me and I was genuinely exhausted afterward!
Going full agentic
If AI can on a weekly cadence tweak my SEO parameters by reading Search Console and PostHog analytics, why can’t we apply a similar continuous optimization pass to our workout program?
I wanted to try the same autonomous loop that I was using for other things such as SEO optimization:
- Read historical data facts
- Read previous invocation notes
- Analyze
- Output a new invocation note and other artifacts
- Update a persistent “memory”
- Peer review all artifacts and decisions with another model
Claude has a great little feature called Claude Routines: They are automatic invocations of Claude on a schedule. This isn’t unique to Claude and all frontier consumer AI offer this: Gemini Spark has scheduled tasks, ChatGPT has schedules, and Grok has automations.
- The data layer was going to be the Hevy API through MCP Nest, together with Strava MCP
- For the brain I wanted to use Notion
- For output artifacts I wanted it to write directly into Hevy
- For peer review I wanted to use Gemini 3.8 flash
For every invocation, once a week:
- The agent pulls my data from Strava and Hevy
- Reads the last invocation’s Notion document that includes any feedback items and reasoning traces
- Reads the workout memory out of Notion (to understand what machines are available, what I can’t due to my herniated disc, etc)
- Analyze all workout data and notes (prescribed workout vs done, user (my) feedback), and output a new artifact into Notion that explains my upcoming lifts, and how weight changed
- Update global memory in Notion
- Ask Gemini to review everything (I created a private MCP connector for this that’s also running on MCP Nest)
- Incorporate Gemini feedback into the final artifact versions
- Create/Update Hevy workout routines through the MCP with the final plan
Every Sunday, Claude does all this, works together with Gemini and presents me with a new shiny Notion doc of what I will have to do the upcoming week, based on my data from the previous week.

It tells me what I did great, what my focus is, and explains things like why it decided to add a new lift in to test exposure.
From Monday onwards, I then go to the gym, open Hevy and click on “Start Routine“ for my 5 day splits.
If I am able to finish all prescribed sets, the agent will automatically progress the lift higher next week.
When I have feedback, like not being able to do something or wanting a different lift, I either type it directly into the comment field in Hevy, or directly into Notion. The next agent invocation will then accommodate my feedback points for the next week’s plan.

My highly personalized coaching team
I’m now on my 5th iteration, and it’s been going great! It’s like having a highly personalized coach that fully understands my weaknesses, my preferences and my goals.
Every week I’m presented with new routines ready to go, and a full doc in plain English that explains every reason behind every change to the system.
If anything is unclear, I just open the Claude Routine that created this week’s doc and I ask. The full reasoning context that created the doc then answers, or makes changes that I request (like swapping Smith Machine with Dumbbell on my request).
This system keeps surprising me with changes to my plan I didn’t expect: Just recently, Claude and Gemini decided to add pull-throughs as exposure test to see if I can do them with my herniated disc, without pain or any kind of weird sensation at low volume. I kept logging all sets, and 2-3 iterations later, this week the agent team decided to add RDL since I passed the hinge test with pull-throughs, to eventually fully fade them out in favor of RDL.
It feels like an evolving personalized system that learns, tries, and reasons to get me to my goals.
As silly as that sounds, even though AI is not a real person, I feel held accountable: I want to finish all lifts because I know that if I don’t, Claude will scold me for it in the next doc.
I see clear progression on my body, my PRs and I honestly feel great. I am back at actually liking to go to the gym!
- Incorporating AI into my workout routine
- Started having autonomous loops around SEO and tweaking my websites, thought why not try to apply it to workouts
- Occasionally uploaded an entire workouts export to Claude or Gemini to see what they have to say about it
- Discovered that Strava has a Claude MCP and Hevy has an API and a MCP server available
- Hooked up the MCP server to MCP Nest
- Plugged everything into Claude and it ripped my workouts apart:
- Too much “junk volume”, no proper progression strategy
- Entire muscle groups were forgotten
- It prescribed a clear system for all my lifts: 1 warmup set, 1 “ramp”, then go directly to 2-3 sets of high volume. That’s it
- Couldn’t believe it really so I aske Gemini to peer review, and besides a few feedback item, it agreed with Claude
- Automating everything
- Weekly routine
- Reads my weekly stats: All my lifts that I did last week, all my runs and activity from Strava
- Reads the prescribed weekly doc and all the feedback that I inputted in it
- Decides what I need to improve in the following week, then:
- Creates a Notion doc for the following week and tells me all my achievements
- Asks other agents (ChatGPT and Gemini) to peer-review the document to catch inconsistencies and hallucinations
- Updates my Hevy routines with new exercises and bumps the weight of stuff I achieved up
- How do I feel about all of this\
- As silly as this sounds: It has been fantastic
- List some of the stat changes:
- Even though it’s AI, I feel it holds me accountable
- It takes away the “thinking” part. I open Hevy and start my prescribed workout, if something isn’t good, I write feedback and the next week it will change
- I see clear progression and what I have to hit for lifts to go up
- I get explanations of everything: I can ask any question to the weekly agent and it explains it to me, or swaps it out