Your coding agent is only 1% LLM
PLUS: Thinking in Platforms is out August 18
Hey there! Welcome to Platform Weekly. Your weekly pour of platform engineering cold brew. Every week, we bring you the best insights, tools and stories from across the community.
Plus…
Thinking in Platforms, the book Kaspar and I have been writing for the past year, is out August 18!
Last chance to have your voice heard in the State of AI in Platform Engineering! Answer the survey and shape one of our biggest research pieces yet!
Your coding agent is only 1% LLM
It is incredibly easy to spin up a next level day 0 AI demo that blows your mind. While everyone, everywhere, all at the same time is failing to scale it to day 100 success. That’s why we’re going into the belly of the beast with our new conference double launch - PlatformCon San Francisco, and a new format entirely AgenticCon, on February 24th and 25, 2027.
If you dissect the source code of one of the most popular coding agents out there, you’ll see that of its roughly half a million lines of code, roughly 1% actually touch the model. The other 99 is deterministic platform work. That’s loading context, routing capabilities, enforcing guardrails, verifying output etc.
So, is it any surprise then that when the majority of organizations still see AI as merely adopting a copilot - most orgs aren’t seeing any ROI? Writing code is maybe 15 to 20% of shipping software in a large enterprise, so a 20% coding speedup nets you around 3% overall. Lame. At the same time, all your competitors have the exact same licenses and same tooling, so even if we take that gain as gospel - at the end of the day, it’s meaningless.
So what then is the durable gain that AI actually provides?
The durable returns live one layer down in the harness, the governance, and the context that your org itself owns.
That layer is what we are dedicating an entire conference to. AgenticCon is happening February 25, 2027 in San Francisco, co-located with the PlatformCon SF live day on February 24. Two days, one stack, one registration. Day 1 covers platform fundamentals and the path from IDP to ADP. Day 2 goes deep on the agent infrastructure layer that every agentic workflow in the enterprise runs on.
So what’s actually on the map?
The whole event is built dissecting and disseminating the best practice in the agent infrastructure layer. Part of that is helping clear up a few things on these three ideas, nested one inside the next:
Platform engineering is the art of building agentic platforms, the whole system many agents run on.
Harness engineering is the art of building one agent’s operating system, a layer inside the platform.
Loop engineering is the art of tuning how work flows through the platform until it produces repeatable outcomes.
They are not synonyms or contradictory concepts, but a matryoshka doll of AI success.
This logic is the same if you’re building an agentic development platform for writing software, or an agentic finance platform that your company’s accounts use every day.
So first zoom out and see the platform. The widest lens. An agentic platform stacks three layers:
Tooling on top. For software development that might be Backstage or GitHub, but remember, we aren’t just talking about IT platforms ;) so marketing might run on a CRM and analytics manager.
Path specs under that. The repeatable actions and agent rules, all either agent-driven and probabilistic, pipeline-driven and deterministic, or hybrid agent plus gate.
Agent infra at the base, shared by everything above it.
No matter the domain from finance, marketing, or software, which we know and love, the agent infrastructure below is the same.
Now let’s zoom in on the harness. One agent’s operating system for a single spin. Context gets assembled, capabilities get resolved (think tools and MCP servers), execution happens in the sandbox where the model reasons, evaluation gates the output.
(As a platform engineer, you build the harness too. And of course. the governance you set wraps it.)
And finally, zoom into the loop. Work triggers, loads its path, assembles context, reasons, calls a tool, hits a gate. Gate fails? It enriches its context and goes again until the outcome ships. Loop engineering is the work of tuning path and platform until loops land repeatable outcomes.
trigger → load path → assemble context → reason → resolve capability → call tool → hit gate
We’re in that beautiful stage of technology where the knowledge and the winning formulas are out there. We just need to bring it all together in one room. So that’s what we’re doing.
Super early bird is open now. See you there!
Quick bites
Highlight of the week
From the community
Want to catch the PlatformCon talks and workshops you missed? We’re adding more every week to the community YouTube channel.
And that’s a wrap on this week! As always, this newsletter is a community project. So if you have anything awesome to share from the cloud-native world, send it my way.
Stay crunchy🥐
Luca