Agents, MCP & Developer Experience at IBM β‘ π₯ Tech Author, Speaker, Entrepreneur π Sharing opinions (βοΈ hello@hackteam.io)
Santa Clara, CA
- Tweets 16,237
- Following 2,093
- Followers 14,230
- Likes 27,794
Build your own Apple HomeView:
1. Buy a $30 charging stand with a speaker
2. Add the iPhone or iPad you already own
3. Spend the difference on literally anything else
Multiple profiles? Most households donβt even need it
3
496
GPT6.1 Sol literally fixed a GitHub run that was failing for weeks and flooding my inbox in <5mins while I was sipping my morning coffee in the sun
413
Beam (new American OSS model) sits between Qwen 3.8 and GLM 5.2 in the benchmarks
need to try this asap with my own evals
1
4
455
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.
1
3
6
539
GitHub's Rust rewrite gave SDK users an in-process runtime instead of requiring a separate Node process. That's a concrete reason to reconsider a language: changing how the software can be embedded, beyond making the same code run faster.
1
274
Every agent goal needs an exit condition.
Before starting a task your agent needs to know:
β What does βdoneβ look like?
β How do I verify I actually got there?
β How much am I allowed to spend trying?
Without that last one, βkeep trying until it worksβ can become a very expensive loop.
Some tasks are impossible because of a missing credential, broken dependency, unavailable API, or simply a bad assumption in the original plan.
You donβt want the agent retrying the same thing 40 different ways, or reverse engineering their way around an unsolvable ask.
The loop I want is much simpler:
Goal β Attempt β Verify
The task is done when verification succeeds, if not try again while thereβs still budget.
If the budget is spent, the agent should stop and report the blocker.
The more autonomous we make agents, the more important this becomes.
A definition of done tells the agent when to succeed and a budget tells it when to give up.
2
223
Coding agents are changing the tradeoffs of choosing Rust.
For years, adopting Rust meant weighing performance and memory safety against practical concerns: learning curve, existing codebases, hiring, and how quickly a team can ship.
Those concerns made sense. Choosing a language has always depended heavily on what your team knows and can be productive in.
But Rust starts to look more interesting as a default when a coding agent is writing much of the code.
The agent can run the compiler, read the errors, and work through fixes. Rustβs type system and borrow checker give it concrete feedback about mistakes that might otherwise survive until runtime. Cargo gives it a consistent way to build and test.
The time developers spend getting comfortable with Rust has been part of its adoption cost. Agents can take on some of that work, while we still benefit from the performance and memory safety.
You still need to understand and review what gets built. Compiling successfully doesnβt mean the business logic is correct.
But with code becoming cheaper to produce, I think weβll increasingly choose languages based on how well we can verify the output. Rust is in an interesting position there.
Have you also started using more Rust in recent years?
2
1
400
Can we please get Codex to continue automatically when quota resets?
Iβm using a goal and it keeps getting stuck when hitting the 5hr quota
2
380
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
5
1
11
571
forget about github activity stats, these are the only numbers that matter now
(yes i sometimes let my agents take weekends off)
983
the urge to use up all your codex tokens before the week ends
(you never know when a new reset hits)
484
marathon on sunday, looks like training paid off
highest vo2 max ever, higher than it was in my 10s and 20s
3
9
431
The default 5min TTL on Anthropicβs prompt cache makes your LLM token bill up to 25% more expensive
You pay for cache writes for context that can only be read from cache for 5minsβ¦
1
1
467
Your agent doesn't need a bigger context window.
It needs you to stop filling it with everything.
We spend a lot of time talking about larger context windows, I think the more interesting problem is learning how to use less of them.
A model can have a 1M token context window, that doesn't mean filling it with 1M tokens is a good idea.
The context window should be working memory, not your documentation dump.
What is your favorite agent harness?
2
1
456
cursor is part of an actual space company, yet openai is making space related commercials for codex
missed opportunity here
piped.video/bJcA23ckzcY?si=e5Sdβ¦
694
Developers love Kimi K3 so much, it had to stop accepting new subscribers.
Much better IMO than reducing quotas for everyone or throttling everyoneβs access to the models.
1
3
428
βLet Kimi Code accept new subscribers againβ
Thanks Tom Cruise for delivering this message during the World Cup final
Kimi K3 has received far more love than we expected, and our GPUs are feeling it.
Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritizing compute for current members. Existing subscribed users are not affected.
We're adding capacity as fast as we can and will reopen new subscription spots in batches.
Going forward, we'll also split membership into two more focused plans: Kimi Membership for Kimi Web, App, and Work; and Kimi Code Membership for coding workflows. This will help us match compute more precisely and keep the experience stable.
Thank you for your patience and understanding!
447