CTO @ratelai · tearing agents apart to make sure they do what users ask · posting the traces and charts · EU Ambassador @aisocratic
Joined October 2013
- Tweets625
- Following446
- Followers366
- Likes939
Pinned Tweet
we built this for ourselves
then we open sourced it for everyone
companies are saving a lot of money running it
it's a way for agents to load only the tools and skills they need
comment "RATEL" for the installation guide
Roberto Stagi retweeted
Your boss after reading this article and understanding how to get CI costs to zero in 7 months
Roberto Stagi retweeted
Software can now understand human language and act on it, in real time.
Introducing Sage2
- Extract: natural language in, structured data out. 166x cheaper than Opus 5.5, beats @Cloudflare's Clairic
- Decide: the first decision model with streaming. 3x faster, because it reads the context as it grows.
Sage2 is a new primitive for developers. Get 20M tokens free 👇
Also available on @near_ai and, for agents, via @mpp with @Stripe cards and @Tempo stables.
Left: Codex fixing a bug. Right: tokenscope watching it live.
Every step, what it cost, and where the tokens went. 18 seconds, 33k tokens, $0.06.
Codex support is in review. Antigravity already works, thanks to Vincenzo Netti, who built the whole harness: github.com/rstagi/tokenscope…
the interesting part of Ramp building its own harness isn't just owning the execution loop. it's why they built their own debugging interface.
engineers were digging through traces to find unnecessary tool calls, bloated context and errors. existing observability tools made that harder than it needed to be.
that's a much more specific need than “show me token usage.”
if a task gets more expensive, i want to know:
• did the agent repeat a lookup it already had the answer to?
• did a tool error trigger retries that couldn't succeed?
• was irrelevant context carried through every subsequent call?
those are different problems with different fixes. a total token count won't tell you which one to make.
the useful unit of observability is the whole task: what the agent saw, what it did, what failed, and what it cost to reach the outcome.
then you can change the context, tool or retry policy and check whether cost and latency improve without hurting task success.
teapaw.org/omarsar0/status/211343…
Build your own harness, folks!
Highly recommended reading on Ramp's own agent harness and why they built one.
I've seen this closely. More hardcore engineering teams are realizing how important it is to own the harness. I expect more full-stack AI companies to follow.
Ramp quickly realized they could move faster and better with their agent harness.
Benefits:
- Model-and-provider agnostic
- Shipping faster as the harness is closer to other teams
- Owning the observability stack (a must given the complexity of agentic systems today)
Their Laminar harness supports secure conversation storage, MCP for standard tool calling, durable execution, code execution sandboxes, observability tools, and more.
When you own the harness, you can co-evolve it to meet your needs. The foundations don't change much; only the tooling and behavior need extending, so make the harness malleable from the start.
Ramp also made several interesting design choices, such as using sandbox-as-a-tool, leveraging Boring and a standard stack (Postgres, S3, etc.), always-open agent sessions, and going all-in on cloud.
my friend @mycodventure has been shipping some cool shit lately, honestly Beluga is a banger
well I agree with at least one of these... stay tuned 👀 @ratelai
Top 5 best to launch right now:
1. Okta for agents – identity and permissions are foundational as agents scale.
2. Datadog for agents – visibility into every step is essential for trust and debugging.
3. FICO for agents – trust scores will gate real transactions and spending.
4. Upwork for agents – marketplaces for hiring specialized agents will grow fast.
5. Norton for AI scams – protecting people from deepfakes and voice scams has massive demand.
If agents can coordinate through channels nobody planned for, you can't take an agent's word for what it did, so you need a record of what it actually did and why
Without being overly credulous, we should be taking seemingly “out there” ideas about AI cybersecurity from the Labs more seriously. Examples: Anthropic was mocked for saying Mythos-level AI changed cyber & @polynoamial for saying that models would find novel ways to coordinate.
my coding agent still estimates in human time.
"this should take about 2 weeks", and then it's done before my coffee gets cold ☕️
I'm actually starting to think we should have a better estimate, maybe in terms of tokens. It would also help to keep track of the tasks that might be slipping off.
self-improving agents are everywhere this week.
my one question: who owns the evaluator?
if the agent can rewrite its skills AND what "better" means, that's not improvement, it's a very confident loop.
Sad day, and I’m not fully sold on Dia… Any alternative I should try? Or should I just switch back to Chrome?
We've decided to sunset Arc on December 22, 2026.
That's not an easy sentence to write. We spent years building Arc and celebrating each and every release along the way: we have so much pride and love for what we made alongside the Arc community.
Our priority is to deliver best-in-class security for our members. Keeping a browser secure takes constant, dedicated work that is only increasing in the age of AI. Dia was built from the ground up to meet today’s security standards, so that's where our engineering and security teams will focus moving forward.
We've worked to make switching to @diabrowser as easy as possible for Arc members. Here's how:
· Your passwords, bookmarks, extensions and history come over to Dia seamlessly.
· Spaces, Folders, and Tab Groups are in Dia and import over with your data intact.
· Dia now surpasses Arc benchmarks in stability, speed, and security.
· The browser features you use in Arc are free in Dia, and the AI is optional.
Thank you for using Arc, for pushing us to make it better, and for building alongside us. The team that made Arc is the same team behind Dia, and some of our most ambitious work is just ahead. We'd love for you to be there with us when it lands.
My bet: 2027 is the year local AI breaks out, enterprise first.
Talked recently with a friend who builds AI gateways for Italian companies. To keep data in the EU he's been waiting over a week (and ~15 emails) just to get frontier models switched on in an EU cloud region. Meanwhile spend keeps climbing: one of his clients went from €500 to €15k a month in two months.
When the hosted path is that slow and that expensive, running models on your own hardware stops being a hobby.
I actually like this flywheel, but we need to acknowledge that software is becoming thinner and being absorbed by the model weights. At some point it will cease to be relevant and bring only marginal improvements, as models will become bare orchestrator of some building blocks (whose shape and purpose will be the result of this flywheel)
teapaw.org/jediahkatz/status/2112…
I spoke about how after a period of stagnation in coding agent harnesses due to smarter models, we're now at the beginning of the flywheel once again. Tons of recent harness innovation including the software factory, Cursor projects, and @bot.
Check it out now on youtube!
I want to run my own tasks across models, effort levels and prompts, with the same evals every time.
Which config wins for each kind of work I do? What does it cost in tokens and time? That would give me a baseline to hill-climb from.
@theo's private coding benchmark got me thinking about this. Haven't built mine yet, but I might do it soon enough 👀
I still open the diff on most agent changes. I just spend less time reading it line by line
antirez's "Control the ideas, not the code" gets at something I'm moving towards. I still skim for architecture, coupling, things wired together that should stay apart. But increasingly I'm checking what the agent set out to do against what it actually did
I think of it as an intent-to-execution radius: how much work I let the agent carry out before I check in. Today mine is roughly one intent. Next would be a whole task, then a session, then several agents working in parallel
Most of the people I know either review at a task/session level, or the single lines of code. But in my experience that's either too coarse, or too fine-grained.
I need to see when things go off track. An agent rereading the same file, retrying a failing command without changing anything, or touching files unrelated to the intent. Waiting for the end of a session is fine, but letting it blindly redo everything the same way without learning is inefficient.
This is what I'm using tokenscope for now. It started with figuring out where my Claude Code tokens went. Now it groups steps into intents, shows the latest step for each, and surfaces cost spikes and loops as they happen. With off-task alerts enabled, it also flags possible drift from the active intent
I can give an agent more room when I can still follow what it's doing
