Agents, MCP & Developer Experience at IBM ⚡ 🥑 Tech Author, Speaker, Entrepreneur 🚀 Sharing opinions (✉️ [email protected])
Santa Clara, CA
Joined November 2016
- Tweets16.2K
- Following2.1K
- Followers14.2K
- Likes27.8K
Roy Derks 🚀 retweeted
I'd much rather work with someone who flexes much they accomplished with the fewest amount of tokens vs how many tokens they spent, and I sense this is a major trend in general
How we quantify this as an industry is still TBD, probably a major new set of benchmarks
Roy Derks 🚀 retweeted
We automated writing code and immediately started writing 800-line instructions about how to write code
Software developers remain undefeated
Roy Derks 🚀 retweeted
Saying software is dead because anyone can build it is like saying photography is dead now that everyone in the world has a camera in their pocket
Not only is it not dead it's the beginning of the Kodak Brownie moment for software, and we will see an absolute explosion of it
It will be the first generation for who building software will feel no different than using it. This is crazy!
The quality and bar of what we expect from our everyday apps will be raised dramatically, we'll have personal software, we'll have disposable software, and agents will be the foundation for most of it
Tech has always been chaotic and evolving, the people who have been able to adapt and move in front of new opportunities have always excelled, this is nothing new, I'm seeing it now and we'll see it 10 years from now
Find the opportunities in the chaose and disruption, there are countless of them
Roy Derks 🚀 retweeted
If you want PMs and designers to build with agents, give them a starter app with your auth, design system, tests and preview deployment already wired up.
Telling people and their agents to follow engineering standards works better when they can start from a working example.
Roy Derks 🚀 retweeted
Stop caring so much about model releases.
Focus on the harness, and improving the environment your agent operates in.
You'll find yourself far ahead of the curve.
Roy Derks 🚀 retweeted
jev is fun to play with but you have to be careful shipping it to prod, over the past few days we accidentally racked up $3.55 in usage
Roy Derks 🚀 retweeted
we built muse from scratch, but it is definitely heavily inspired as a product by openclaw. After I used openclaw in january I bought hundreds of mac minis for the MSL team and lots of us fell in love with using openclaw (and other personal agents). @steipete is a genius and his harness was pioneering from the jump. I think a lot of people were inspired by it. our goal with muse was to build something like openclaw that we could make safe and secure and easy to use and scale to billions of people.
Roy Derks 🚀 retweeted
When a multi-agent demo finds more bugs, check whether it also searched more code and spent more tokens. Compare the same scope and budget before crediting the architecture. A bigger search can produce a bigger result without being more efficient.
Roy Derks 🚀 retweeted
total MCP victory, some quick misc thoughts about why MCP is so much better than CLIs:
- indexable tool catalog letting agents scale to unlimited tools
- no requirements to be running a full sandbox
- consistent auth across all MCPs rather than each CLI inventing its own auth
- multi account support for all MCPs unlike CLIs
- implementations like code mode let the model know what will be returned allowing for super efficient token usage unlike CLIs
the reasons it took this long for MCPs to finally have their moment is mostly due to bad MCP implementations in clients:
- you had to restart your whole client to use an mcp (no hot reloading)
- agents weren't as familiar with debugging mcps as they were CLIs, so it was a lot easier for people to get set up using them
but over the past year, things like codex and claude plugins have all been using MCP under the hood, i.e computer use is an MCP, i believe claude artifacts are an MCP app, just the silent steady adoption
there is still an element of MCPs that is 'this MCP could've been an OpenAPI spec' but that'll go away as things like triggers get more adoption
at the end of the day, what's important to realize is while yes there are these difference between CLIs / MCPs / etc they're all just different ways of doing tool calling, and you can do some combination of lazy loading, searchable tools, and filtering to build efficient harnesses
Roy Derks 🚀 retweeted
I hate to say it, but if you’re running more than 2 sub agents at time, you’re almost certainly burning tokens for 0 quality gain.
Agents don’t trust each other enough to avoid double-checking everyone’s homework.
Roy Derks 🚀 retweeted
memory has been a hot topic for the past ~2 years
every time we bring it up or do anything there, gets ton of interest, but nothing sticks
few takes on why below:
- how to update and then use memory needs to be pretty tightly integrated with the harness. so its hard to build a standalone memory product
- the hardest part of memory is the logic of deciding what to remember. storing and even querying memory is not necessarily hard. the logic of deciding what to remember is often (a) application specific, (b) just prompting, which makes it hard to productize
- memory is not that useful in a general purpose agent, but more for when you are doing repeated tasks. more high usage ai agent are not just doing the same task repeatedly
- memory hasnt really proven to be useful for coding agents, and most agent things that take off happen first in coding agents
Roy Derks 🚀 retweeted
AGENT HARNESSES ARE THE NEW GPT WRAPPERS
What exactly is an agent harness? A harness does 4 things:
1. Runs the model in a loop so it keeps working step after step instead of answering once and stopping.
2. Gives it hands to read files, call tools, open portals, and run code.
3. Manages its memory so hour three of a job still knows what happened on hour one.
4. Enforces the rules about what it can touch and when it has to stop and ask a human.
WHY IT'S INTERESTING
- The market is MEGA. I think about it like a wrapper let you sell software ($800B market) but a harness lets you sell the work ($5T+ market)!
- It doesn't depend on any one company's model. The knowledge about the job lives in the harness, so you can run GPT today, Claude next month, an open model like Google Gemma, Qwen, Deepseek etc on your own machine after that, and it keeps working. A wrapper was one model doing everything and a harness is a router.
- It gets better the more you use it. See, every correction a human makes becomes a rule the harness keeps, so what you have after 500 jobs is a completely different product than what you had after 5. Wrappers only got better when OpenAI got better.
- You can charge for finished work! The harness knows when a job is done, so you can price per claim, per filing, per review, per resolved exception, or per closed month. This helps compete with SaaS!
- Most jobs are just the same handful of decisions, repeated, using the same few tools. That's true for most of the 800+ occupations out there, which is why almost all of them could have a harness built for them. Lots of space for startups to be building.
-The frontier labs won't come for these. OpenAI is not going to learn how a freight claim gets denied in Rotterdam or which prior auth your specific payer rejects. Those markets are too small for them and the knowledge only exists inside the job. They'll keep making the models better, which just makes your harness better which is cool.
- OH, AND really important, you can ACTUALLY build one now. OpenAI's Agents API, Claude's managed agents, Vercel's eve, LangChain harness framework etc. They all make it way easier to build agent harnesses.
Agent harnesses truly are the new GPT wrappers.
This is just the beginning.
Roy Derks 🚀 retweeted
Why run local models yourself?
LARRI is a simple open source CLI that rents a GPU, brings up an open weights model on it (Qwen, GLM, Devstral, whatever fits) and tunnels it to a fixed local endpoint
This looks pretty cool, and with commands to spin up and down solves for maintaining the infra yourself
Roy Derks 🚀 retweeted
If a model post-trained on Kimi K3 can beat Fable/Astra we’ll be in for a lot of fun
Roy Derks 🚀 retweeted
A massive difference between Claude Code and Codex:
- Codex lets you use any model you want (not just OpenAI) and harness is open source
- Claude Code doesn’t and is closed source
Given this is the two leading AI labs, notable difference in approaches
Codex doesn't lock you into OpenAI models. Tibo @thsottiaux on why:
"If you are part of this community and building an excellent coding harness, why would you couple it to your model? That felt quite disappointing, to make that decision. It didn't feel right.
And it's open source in the first place. It would have been trivial for anyone to fork it. But then you're encouraging people to go and use that fork, and now suddenly you have overhead and the only reason you have a fork is because you wanted to change 10 lines of code to add support for another model provider. That feels very silly. So why not just support it in the first place?
The other thing is we benefit a lot from being able to give optionality. Maybe today you love using OpenAI models and you're super productive with them, but tomorrow there's a new model that comes out and you want to try that. Why force you to completely change your setup just to try a new model?
Optionality is very important to companies that we work with. And this is something that we absolutely lean into."
