Cesar, algo mas (@zerojuls) header
Cesar, algo mas (@zerojuls) profile picture

Si decides, sígueme ... pero no se a donde voy ...

Lima - Peru
Joined February 2009
Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)Photo by Cesar, algo mas (@zerojuls)
Cesar, algo mas retweeted
Héctor Camacho: “When he left Chulucanas, I was truly very sad. He was still so young, and I thought, ‘Now we’re losing Father Roberto.’ I assumed that, at some point, he would surely give up the priesthood, fall in love, and get married. That's why it impressed me so much... 1/6
3
212
14
1,284
52,776
Cesar, algo mas retweeted
Los vecinos de la plaza Butters tenemos que tolerar los gritos de un culto religioso en plena vía pública. He consultado al Serenazgo y me dicen que el municipio, al mando de Jessica Vargas (Renovación Popular), les ha dado permiso.
86
235
15
1,032
47,785
Cesar, algo mas retweeted
Everyone is talking about software factories. "Factories, factories, factories." Nobody wants to talk about the unsexy part: the checks that tell you whether anything the factory built is actually right. Here are the steps that make a factory safe to run on code your customers depend on:
11
11
2
39
6,174
Cesar, algo mas retweeted
a16z just opened a fellowship for one of the hottest jobs in AI. Forward Deployed Engineers. The people who take AI from demo to working inside real companies. Cohort 2's founding fellows include FDE and Applied AI leads from: - OpenAI - Anthropic - Perplexity - Replit - Harvey - Cognition - Exa What it is: - 8-week cohort starting mid-January 2027 - Dinners and events every other week in SF and NYC - You can be based anywhere Who it's for: - FDEs and Applied AI leaders building and deploying AI with customers Apply by Dec 31. Reviewed on a rolling basis, and spots are limited. build.a16z.com/fellowships/f… Bookmark it and send it to the FDE friend who should be in that room.
11
25
4
430
68,418
Cesar, algo mas retweeted
29
228
5
2,684
387,763
Cesar, algo mas retweeted
After two years of rapid change in software development, we are again at a crossroads, with two competing visions for what the future might hold. Vision one is software factories (e.g. Warp Factories), code-defined, closed-loop systems for continuously measuring and improving cloud coding agents. Software factories are developer infrastructure. Platform teams do “factory engineering” on them to optimize development across an entire product, getting the most productivity from agents while minimizing cost and improving quality. It’s a system design mindset for deploying agents. Vision two is agent teams (e.g. Devin Bots). In this approach, individuals manage agents to orchestrate complex tasks. For example, you might have a “chief of staff” agent and ask that agent to do a coding project. It decomposes the project into subtasks that different agents take on (e.g. a designer, an implementer), reports on progress, and eventually ships. These teams can also have “always on” agents that proactively fix things, respond to events, and so on. Unlike with software factories, the user experience is more about “managing” agents, akin to how you might manage a person or a team. Which one is better depends on how you conceive of the problem. I view the problem as: how do you securely deploy coding agents to maximize product throughput and quality, while minimizing cost? I believe this is what businesses care about – there is this magical new technology for building software at scale and they want to make sure they are using it to get the biggest possible competitive advantage. Put simply: this is an optimization problem. We have a new engineering primitive, the coding agent, and we want to use that primitive as well as we can. Framing it in terms of optimization immediately leads to a need for measurement; after all, you can only optimize what you can measure. In order to get to measurement, you need reliable, standardized data. In order to get data, you need some amount of centralization and standardization, otherwise that data would be too noisy and heterogenous to do anything with. In order to centralize, you need to move to the cloud and define what a factory is in code. Your factory needs to be open and flexible so you have degrees of freedom for testing different models, harnesses and skills. You need ways of experimenting to see if different configurations perform better. And then you need to be able to feed those learnings back into the system, so it improves over time. When you frame the problem clearly, the shape of the solution becomes much more obvious. Compared to agent teams, the factory engineering approach is much better suited to solving the optimization problem because it puts measurement, repeatability, and standardization at the center. It allows for a principled approach to improvement in the long run that is missing with agent teams. You want your factory to scale like an engineered system, not like an increasingly large organization (don’t make your agents work in a big company!). The engineering approach is possible precisely because agents are not people; for the first time in the history of software development we can directly measure, grade and improve how code and products are shipped. We shouldn’t waste this opportunity by anthropomorphizing the technology and treating it like a teammate. Also, wouldn’t you rather engineer an agent system than manage a team of agents? I’ve been both a builder and a manager for most of my career. You can guess which I prefer.
74
31
21
475
48,961
Cesar, algo mas retweeted
What you read on Twitter (individual engineers using Claude Code) is very different than what moves the needle for large companies How to redesign enterprise software development lifecycles, you need to give this a read:
I've been working with a few orgs with 700+ engineers to redesign their sdlc. they are confused. they've rolled out claude etc., engs are more productive, but they're not seeing almost any of the "speed" startups are feeling right now. i keep telling them, the software part is easy(er). you've got to reorganize people. read some research today that I think nicely illustrates this. SWE-CoWork (will drop link) found that frontier models completed *less than 8%* of their assignments when they had to track down requirements spread across a 26-person org of AI-simulated colleagues. Anyone working with an agent day to day can tell you that an agent doing less than 8% of anything is virtually impossible. but also that asking 26 people for requirements is insane (but common). agent improved when given all requirements up front, but still struggled on changes that touched 60+ files. so, some of the speed will be realized when models get better at highly distributed implementations. big labs will do that for you. but the real key to moving faster is redistributing ownership to be narrow and deep, only then followed by said "software factory."
9
12
306
28,671
Cesar, algo mas retweeted
𝗠𝗖𝗣 𝘃𝘀 𝗔𝗣𝗜. An 𝗔𝗣𝗜 defines how software systems communicate through specific endpoints, requests, and responses. It gives applications a structured way to access data or trigger functionality in another system. 𝗠𝗖𝗣 gives AI applications a standardized way to discover and use external tools, data, and resources. Instead of building custom integrations for every AI client, an MCP server exposes capabilities through a common protocol. MCP doesn't replace APIs. Most MCP servers wrap them. APIs expose the service. MCP makes it usable by a model. Or a human. The barrier to entry for non-technical people to interact with platforms has been drastically reduced. But once a model decides what to call, the work it does gets harder to track. Agents finish tasks without status updates while people shift priorities. Permissions limit what an agent can touch, but they don't show whether its work still matches the strategy that funded it. Tempo Loop connects to the tools teams already use, like Jira, to 𝘁𝗿𝗮𝗰𝗸 𝗮𝗻𝗱 𝗴𝗼𝘃𝗲𝗿𝗻 𝗵𝘂𝗺𝗮𝗻 𝗮𝗻𝗱 𝗔𝗜 𝘄𝗼𝗿𝗸, 𝗮𝗻𝗱 𝘄𝗵𝗮𝘁 𝗶𝘁 𝗰𝗼𝘀𝘁𝘀, 𝗶𝗻 𝗼𝗻𝗲 𝗹𝗶𝘃𝗲 𝘃𝗶𝗲𝘄. When work drifts, it recommends a fix while there's still time to act, then updates those tools once the change is approved. 𝗦𝗲𝗲 𝗵𝗼𝘄 𝗧𝗲𝗺𝗽𝗼 𝗟𝗼𝗼𝗽 𝘄𝗼𝗿𝗸𝘀 → lucode.co/tempo-loop-z73xd What else would you add? —— 💾 Save it for later. 🙏 Thanks to @TempoHQ for sponsoring this post.
15
205
5
1,149
63,247
Cesar, algo mas retweeted
Few days back, I published LangChain 101: a 387-page, completely free, open-source manual. Most LangChain tutorials show you how to create an agent in a few lines of code. But what actually happens underneath? What happens when an agent crashes, retries a tool, loses its state, or needs to resume? I wanted to answer those questions with code and real measurements. So I went through the source code of LangChain 1.5, LangGraph 1.2, Deep Agents 0.7, and LangSmith, built reproducible experiments, and documented the results. A few interesting findings: → Does LangChain improve agent accuracy? A plain tool loop, create_agent, and a hand-built StateGraph produced identical outcomes across 60 shared local benchmark tasks. → Deep Agents has a token cost. In our tests, it showed surprising results with prompt tokens than a plain loop. → Checkpointing doesn't guarantee exactly-once execution. Tools can run again after a crash. We tested the failure scenarios and how to prevent duplicate side effects. → Structured output isn't equally reliable across providers. We tested multiple strategies against local and hosted models and documented where they fail. The manual goes much further: 1️⃣ Agent internals, middleware, and tool execution 2️⃣ LangGraph state, memory, streaming, and durability 3️⃣ Human-in-the-loop, interrupts, and approvals 4️⃣ MCP, Deep Agents, LangSmith tracing, and evaluations 5️⃣ Production architecture with FastAPI, PostgreSQL, observability, and failure recovery 387 pages. Real code. Benchmarks. Architecture diagrams. MIT licensed. No black boxes. Every important claim is backed by source code or a recorded experiment. 📖 Read the full manual (free): aiengineeringfromscratch.com… If you're building AI agents with LangChain or LangGraph, I hope this saves you a lot of debugging time. Next, I'll be working on browser agents!!
🤖 Made with AI
4
22
1
182
5,712
Cesar, algo mas retweeted
Is grep all you need? This work claims that an append-only text log plus a regex search tool beats summarization and structured memory for long-horizon agents. (bookmark it) GrepMem keeps the agent's full history as plain text and lets it search that log when it needs something. It does not need embeddings, a retrieval index, or an extra model. They trained Qwen3.5-9B with RL under five context strategies. GrepMem reaches 95-97% on ContextBench, where lossy methods reach at most 21%. It beats SUPO summarization by 8.8 points on BCP-Multi and stays on the accuracy-and-token-cost Pareto frontier on all five tasks. Why does it hold up? Summaries decide what to keep before the agent knows what it will need later. The log keeps everything, and search finds the detail on demand. arxiv.org/abs/2611.25123 Chat with Paper: academy.dair.ai/papers/forge…
25
22
8
204
11,063
Cesar, algo mas retweeted
La reacción de Curwen al Temblor ☠️⚠️
TEMBLOR LA PUTA MADRE!! ESTOY TEMBLANDO
106
405
54
7,448
310,114
julian casablancas es el idolo de los violines, alcoholicos y manipuladores, the strokes es musica para nenazos que se creen rockeros por escuchar una guitarra y poco más!! en fin nos vemos el dia de hoy en el estadio de san marcos a las 7pm cancha 2
19
169
19
1,483
25,990
Cesar, algo mas retweeted
read of the day: fs.blog/first-principles
4
43
1
485
31,736
Cesar, algo mas retweeted
benchmarkmaxxing is the new tokenmaxxing
We spend 2 trillion+ LLM tokens a month. That’s way too much money spent to trust our agents based on gut feeling. So, we started using custom benchmarks. We ran into plenty of problems, from parallel runs to open-ended tasks. Here’s everything we learned along the way:
1
5
216
I asked people to break my website live at the @aiDotEngineer workshop yesterday, by sneaking hacks through @Posthog Self-driving. Top 3 attempts: - Hidden custom form to query any SQL table without admin access, mixed in with a legit feature request (Bobby Tables). - Backdoor "debug" mode to enable when a specific input is provided, which would show data on currently logged-in users (Mr Banana). - Shut down the website because of a fabricated security issue (CVE, anonymous). None of them, happily, passed through, as they were either blocked by Self-driving safety filters or marked as non-actionable. Gonna do it again 100%.
1
2
1
7
644
Cesar, algo mas retweeted
imagine a software factory discovery + implementation layer that doesn't get popped by security exploits that's posthog self-driving
I asked people to break my website live at the @aiDotEngineer workshop yesterday, by sneaking hacks through @Posthog Self-driving. Top 3 attempts: - Hidden custom form to query any SQL table without admin access, mixed in with a legit feature request (Bobby Tables). - Backdoor "debug" mode to enable when a specific input is provided, which would show data on currently logged-in users (Mr Banana). - Shut down the website because of a fabricated security issue (CVE, anonymous). None of them, happily, passed through, as they were either blocked by Self-driving safety filters or marked as non-actionable. Gonna do it again 100%.
2
2
6
472
Cesar, algo mas retweeted
In many software engineering environments, there’s nothing wrong with being in a state of partial understanding. In fact, in large systems a partial understanding is the best you can do.
17
27
7
425
66,022
Cesar, algo mas retweeted
man, this is so good. i tried to argue something similar in dylanamartin.com/2026/10/13/… but sean's points are at lot more specific and salient than mine, which on second reading feel pretty general
In many software engineering environments, there’s nothing wrong with being in a state of partial understanding. In fact, in large systems a partial understanding is the best you can do.
1
1
5
234
Cesar, algo mas retweeted
Moving from "traditional" product engineering to AI product engineering wasn't the leap @dmarticus expected, but he still had to learn new things: writing evals, balancing human vs agent interfaces and more. dylanamartin.com/2026/11/19/…
3
13
1
122
6,240