Each week, TTY Lunch brings together exceptional builders around the table, this time under the sun on the Tech Makers rooftop in Amsterdam. Today’s lineup included Alina Dumitrache and Fabien Bouhier (Tech Makers), Guillaume Lesur (Wire), Justin Halsall (RRweb), Sohrab Hosseini (Orq.ai), Robert Hommes and Vladimir Vučković (Moyai), Wilco Kruijer (Rejot), and Zac Allen (Zero Servo).
Red-Green-Refactor with Agents
Justin Halsall detailed a workflow that repurposes the Red-Green-Refactor cycle for an agentic world. In the Red stage, he tasks the agent to write one failing test case and verifies that it is actually failing. In the Green stage, he directs the agent to make that failing test pass. As he summarized, “you write the test in a state where it’s failing, and you verify that it’s failing. And then you go and fix it.” Vladimir Vučković and Zac Allen noted that starting from failure tends to produce better outcomes than generating code first and testing afterward, even though this workflow takes more preparation and well-defined intent.
Useless Memory, Thin Harness, Fat skills
A critical distinction emerged between agent memory and “skill” creation. Robert Hommes expressed skepticism toward agent-generated memory, arguing that when models write and update their own memory, they can reinforce their own failure modes rather than correct them. It has been shown that curated context (memory or skills), written and maintained by humans, is much more effective than dense, agent-generated context. But this requires discipline: no memory is often better than poorly written human memory. Wilco Kruijer described attempts to manage context externally, for example by storing structured notes and reloading them into sessions, but these approaches were seen as brittle in practice. In contrast, the group discussed building “skills” as more reliable: encapsulated behaviors that can be tested, evaluated, and iterated on. Robert explicitly framed this approach as “thin harness, fat skills,” emphasizing that the surrounding infrastructure should remain minimal while the core capabilities are carefully defined and improved. He described a workflow where skills are created, tested across multiple cases, and evaluated using labeled outputs, with strong examples reinforced over time.
MCP + Skills, The Distribution of Roles
The group explored the division of labor between MCP and specialized skills. Sohrab Hosseini described how Orq.ai utilizes MCP as a standardized connectivity layer, exposing tools and data sources. In this framework, skills sit on top and define how those tools are orchestrated to complete tasks. As he explained, MCP provides primitives, while skills guide the sequence of actions needed to achieve a goal. Guillaume Lesur added that even with this structure, a key challenge remains in how the model decides when to use these tools, noting that agents do not always trigger the right behavior or look for the right information to guide their decisions. In that sense, while MCP and skills define the available primitives and workflows, they do not by themselves ensure that the agent will apply them correctly.
The Unsolved Frontier For Coding Agents
The group shared issues that remain unresolved in agent workflows. A recurring challenge is that agents often reason locally, looking at a limited subset of files and reproducing patterns without understanding the broader architecture. This leads to code that appears correct but is inconsistent with the overall system design. There was also frustration with agent behavior such as guessing instead of searching, using outdated packages or APIs, and being overly defensive, for example maintaining backward compatibility where it is unnecessary. Participants also noted that increased speed can lead to feature sprawl and additional cleanup work, reducing the net productivity gains. Overall, agents still require close supervision to ensure correctness and coherence.
Finding the Right Co-founder
The closing segment addressed co-founder dynamics. Participants discussed the difficulty of finding the right partner, especially when there is an imbalance around ownership of the idea or motivation. One observation was that joining a project with a predefined idea can make it harder for a co-founder to feel equal ownership. There was also discussion of mismatches between expectations and reality, particularly when people are attracted to the idea of a startup but not the execution, aka the “grind”. A common suggestion was to spend time working closely together before committing, in order to test compatibility under real conditions.