Patrick Debois

A diagram showcasing the evolution from LLM to Agents and beyond - Does this align with your journey ? #aiengineer

llm= road, coding agent = car, pilot = coder, context = fuel . What else would you add to the metaphor?

So I hear Claude Code will charge extra for "-p". Here's a simple gist to do a headless call without it. gist.github.com/jedi4ever/75be…

The best AI coding conference is coming to London 🎉 and I have a 50% discount code for you (PATRICK50). AI Native DevCon, June 1-2. As content curator, proudly biased, I genuinely think we've built the best program on AI coding out , see you there ! - ainativedevcon.io - luma.com/aidevcon-ldn26…

@buildsafter5 @swyx here here

@patrickdebois @swyx sure, but knowing what failure looks like still beats not knowing. even if you can't fix the rate.

this one is doing v well btw if you want the popular vote filter on the firehose of all the things @patrickdebois was one of the track keynotes i gave a "blank check" to based on his sincere support since our very earliest days + when in europe we must feature the DevOps guy. he didnt disappoint!

@vidimitrov true !

it’s not that they want to write docs, we both know that docs drift fast… what they really want is to: - stop repeating themselves (reconstruct intent each new session) - talk to an agent that is fluent in their context and recalls historical decisions - get consistent results that belong with their codebase (conventions) - trust the outcome (verification procedures) - keep only persistent context that cannot be recreated from reading the code - share the same experience with their team the packages approach is good but it doesn’t cover the full spectrum of context types

agree on the "rigor", but "context is the new code" isn't quite right what @patrickdebois describes was always part of code creation - lived in our heads and scattered docs, just never captured its not new - we need the right taxonomy to define it and a way to materialise it reliably

@buildsafter5 @swyx agree on the tdd. But the non-determism makes it feel more like error-budget and risk mitigation vs full testing

@swyx @patrickdebois TDD for context is the right framing. you see the failures long before you understand them.

@jvetterle @aiDotEngineer and rots and needs maintenance, so there is friction!

@aiDotEngineer @patrickdebois basically context is the only AI dev artifact that compounds. Models commoditize, workflows shift, frameworks rot but the context graph (rules, examples, constraints, learned-from-failure) gets more valuable over time.. everything else is rented.

Context may be the most under-engineered layer in AI coding today. In this keynote, @patrickdebois, argues that if agents are driven by prompts, rules, and memory, then context deserves the same rigor we already give code. youtube.com/watch?v=bSG9wU…

@ai_carmichael @aiDotEngineer it’s real! I hear it over and over again. We need better tooling and processes around it.

@aiDotEngineer @patrickdebois Agreed. the gap I keep hitting is that context has no ownership model: no versioning, no diff, no rollback when an agent starts drifting because something upstream changed silently.

@fatihaydindev code > context > knowledge indeed !

thoughts on context management. Better knowledge = better context management. I'm calling this the triadic hermeneutic loop, building on the classical hermeneutic circle from philosophy (Schleiermacher, Gadamer) and extending it with a third actor: the AI. The main problem isn't model capability; it's informatics. We need structured know-how, shared context, and a maintainable lifecycle [@patrickdebois 's CDLC]. Humans in the loop, i mean, the actual agent (in the philosophical sense) is the human: the one who learns new things, reflects them back into the context, and externalizes knowledge into shared structure. The AI agent executes; the human agent understands.

@swyx Appreciate the trust you put in me. Keep pushing the industry!

Background Agents Summit schedule looking 🔥 The first event of it's kind to explore what agents throughout the SDLC look like. Presented by the companies actually doing it today. @patrickdebois @pelikhan @_colemurray @leodido @csweichel @MattJamesBoyle @suhailpatel @casualzach_

The Human Harness engineer = The person that enables the Agentic Coder

Wonder if anyone created a website where agents do all the writing and you just feed/curate relevant articles as context.

I created 'Self driving Repo' (. com) - let's you spawn recurring agents with budget control - agents have full FS access and sources like other repos, issues, and the web - agent outputs are pushed to your preferred repo/branch This makes it super easy to setup durable knowledge-base automations in the cloud reply 'yes' for early access

@janwilmake I definitely like the idea and it can work the more refined the harness is around it.

Do you think self-driving repos will become a thing? Most codebases now evolve iteratively by humans prompting them. When will we see codebases that evolve based on a problem in the market, a connection with your notion, and possibly, an ai having conversations with the team? x.com/janwilmake/sta…

Context Coders & Agentic Coders: I see two distinct coder types emerge: the Context coders manage the context development lifecycle and use the agentic pipelines the Agentic coder maintain: a platform provided to other coders. So some will keep building the thing and others build the thing that builds the thing. The latter might be the new CI/CD emerging.

@JamieMallers here here - unlocking organisational flow is key

@patrickdebois The org dimension is the silent killer. Teams invest in better tools and richer context files, but if tribal knowledge stays in Slack threads and postmortems nobody reads, the agent operates with a fraction of what it needs. Context maturity has to be a team sport.

Context maturity for coding agents. Not agent maturity. Context maturity. Three dimensions: 1/your tools, 2/your context, 3/your org. Most teams are advancing one and ignoring the other two. He're the full article - tessl.co/ggm