AI Coding Agents Evolve to Trusted Collaborative Partners

Key Quotes

"For coding we thought okay well what is the best personality for a coder for a pair programmer for somebody who you trust and how do we like eval against that how do we come up with behavioral characteristics and we came up with things like communication it needs to keep you in breast of what's going on while it's working uh planning like come up with a strategy do some searching around like figure out context gather figure out what to do before you just dive in if if if it makes sense too and then you know check your work right and so these are just best software engineering practices that turn out to be behavioral characteristics and we can measure the model's performance on those behaviors and grade it that way"

Brian Fioca and Bill Chen explain that training AI coding agents involves developing specific "personalities" that align with desirable software engineering practices. These characteristics, such as communication, planning, and self-checking, are not just abstract concepts but measurable behaviors that can be evaluated and graded. This approach aims to build trust with developers by ensuring the AI acts predictably and collaboratively.

"so codex is the frontier coding model that we have that is optimized for its harness the codex team is is very focused on creating a coding agent and they wanted to work perfectly inside of the shape of the harness and api that we have so they're completely unbounded it's open source so yes that's open source and the model is available on the api so so that's what they focus on"

Brian Fioca and Bill Chen clarify that the Codex model is specifically designed and optimized for its dedicated "harness" and API. This focus means it's tailored to function optimally within that particular environment, distinguishing it from more general models. They emphasize that this specialized model is open-source and accessible via API, highlighting its intended use for agent-focused coding applications.

"codex loves rip grip so if you make a rip grip tool and tell it to use it it'll use it so if you call it gret it actually does a little bit worse but if you call it rg it actually does really well"

Brian Fioca and Bill Chen share an observation about Codex's tool usage, noting its strong preference for "rip grip" (rg) over "grep." This indicates that the model develops specific habits based on its training data and naming conventions. The authors explain that calling the tool "rg" leads to significantly better performance than calling it "gret," demonstrating how even minor naming details can impact an AI's efficiency.

"we found our customers have found that people really want to follow along with what it's doing so they can like interject or stop it or at least understand what it's thinking so they don't waste all the kinds of time like doing a rollout that they have to throw away so with the five series because it's more general and it's just about as good as coding as codex for a lot of things we've taught it to be more communicative and so it has preambles before tool calls it'll say things like i'm about to go look for this yeah and you can steer that really well"

Brian Fioca and Bill Chen discuss the importance of transparency and control for users interacting with AI agents. They explain that the GPT-5 series, being more general, has been trained to be more communicative, providing preambles before executing tasks. This allows users to follow the AI's thought process, interject if necessary, and avoid wasted effort on incorrect or undesirable outcomes.

"the trend that we're sort of seeing as the abstraction layer really moving upwards from the model layer towards the agent layer as i said we're training our models starting to be a little bit more opinionated especially with regards to going model like codex and then the models are really good at doing certain things one inside of a certain harness type of thing and so we're hacking that packaging that up more closely so we're actually shipping this entirety entire agent altogether than you can actually build on top of that agent"

Brian Fioca and Bill Chen describe a significant trend in AI development: the shift in abstraction from individual models to integrated agents. They explain that models like Codex are being trained with more specific "opinions" and optimized for particular "harnesses," allowing them to be packaged as complete agents. This approach enables developers to build on top of these agents rather than managing individual model complexities.

"i really want to see the the trust level go of even further right like at openai i get to work with some of the most amazing developers i've ever worked with in my life they're incredible like some crazy tech leads i wish every company no matter whether they're like a small dev shop in alaska where i worked for a while or or open ai be able to have on their team like capabilities that you would only be able to get at like a top tier firm right so like so all of my teammates at all of these places could turn to a coding model and be like hey how do we do this like crazy awful refactor that we have to do to get to support this new customer that we have or like wow there's so much of a mess here or like what's the best way to actually implement this new technology and have it be so trusted and so right and so smart that like you know we can actually perform better than we could normally get access to"

Brian Fioca expresses a vision for the future where AI coding agents are trusted to the extent that any company, regardless of size or location, can access top-tier development capabilities. He desires for these agents to be intelligent and reliable enough to handle complex tasks like difficult refactors or implementing new technologies. This would democratize access to advanced developer skills, enabling all teams to perform at a higher level.