Hey there!
Summary of this email:
- What is RPI: Research, Plan, Implement (+ RPI course 100% published)
- 🗞️ Is this the end of coding agents in the terminal?
- Latest Café con Codely recap (check it out here)
- The joke of the week (eclipse-themed!!!)
🔁 What is RPI: Research, Plan, Implement
Research, Plan, Implement are three phases to implement SDD in a simple way:
- 🔍 Research: the agent investigates your code in read-only mode. It looks for the relevant files and modules. It identifies patterns and conventions. It writes everything down in a
research.md. - 📋 Plan: with that research, it splits the task into ordered and verifiable steps. It does not write code yet. The result is a
plan.mdyou can review. - 🛠️ Implement: it executes the plan step by step. It runs the tests (verification) after each step. If something does not fit, it flags it so it can be iterated. That is why RPI is a loop.
The plan is the most important part since fixing a plan is much cheaper than fixing code. Sometimes the agent implements something that does not respect your abstractions. Steering it back costs more than having detected that difference in a previous plan and having made the changes.
That said, the native plan modes of the tools don't convince us because they usually phase horizontally by layers: backend first, then frontend, then tests. That prevents you from shipping each phase to production.
That's why we built our own plan creation skill that uses vertical slicing. You have it in our Agent Harness.
We also believe another thing: that RPI has an expiration date. The better the models and the harness are, the more RPI phases you can skip. That said, nowadays you can still notice the difference between having a plan and not having one for medium or large tasks.
When each phase makes sense, how to make better plans, and real cases of how we apply it at Codely is what we cover in the Spec-Driven Development with Research, Plan, Implement course that we've just published at 100%.
🗞️ News of the week: Is this the end of coding agents in the terminal? (Claude Code, OpenCode…)
Grok Bot was announced on Tuesday. It may look like yet another tool that will go unnoticed, but after analyzing it, we believe it can set a precedent.
It's like OpenClaw, but for everyone.
At the beginning of the year, when OpenClaw went viral, the need for personal agents to do tasks outside of coding became clear.
Then came Hermes, which reinforced that hypothesis.
Grok Bot, the first SpaceXAI product after the Cursor acquisition, takes it one step further, since it has achieved it with a very simple interface: a chat just like Telegram or WhatsApp.
The bad thing about OpenClaw or Hermes is that you have to configure them so they can talk to you on Telegram. With Grok Bot, that interface comes out of the box, along with the ease of adding connectors.
The other difference of Grok Bot is that they don't focus it so much on a personal agent, but on "coworkers". You create bots with roles. For example:
- Accounting manager: It keeps the company's books.
- Inbox manager: It reads your email and warns you if there is something interesting or if everything is spam.
- Expense manager: It reports the company's expenses and any deviations.
The good thing is that you can also create a "boss" role that supervises the others and coordinates them for you.
And the other advantage is that it learns from you. It can watch how you use your browser, for example to work with your day-to-day CRM, and then replicate it. All with a very simple interface.
With all this, it seems we are moving further away from the more niche tools that live in the terminal, towards apps or the cloud directly.
We'll see the impact of Grok Bot. Even if it's not as revolutionary as OpenClaw was when it came out, it can strongly influence the tools that Anthropic/OpenAI/Google create or the improvement of the ones they already have, like Claude Cowork.
Tomorrow we will be live at 9 CEST discussing this story and many more on Café con Codely. On our YouTube, Twitch and X.
☕ Café con Codely recap
This is the top 3 of headlines we discussed in the latest Café:
- Shai-Hulud returns to npm: keyv and flat-cache compromised, and cooldowns are your best defense. flat-cache alone has more than 500 million weekly downloads, and the worm enters through preinstall scripts stealing your tokens.
- Tech job market in Spain: 6,200 offers in July and 93% still do not publish the salary. The European directive requires disclosing the range if you ask before the first interview since June 7, although Spain has not transposed it yet.
- Next.js 16.3: ships next/root-params and ends locale prop drilling. You can read root URL params from any Server Component, and libraries such as next-intl already adopted it.
You have the full Café recap with links to each story at the moment we discussed it.
Tomorrow we will be live at 9 CEST discussing this week's news on Café con Codely. On our YouTube, Twitch and X. 🙌
And since you've made it to this part of the newsletter, here's the joke of the week, which I know you were waiting for:
Yesterday was the day for people who code in Java… Everyone was watching the Eclipse! 😂 😂 😂
Cheers!