Codespeed @codespeedai

We're still in the pre-history era of AI software engineering, and prompting has only scratched the surface. As the cybersecurity shock this week has shown us, the future is far weirder and more expansive than we can imagine. codespeed.ai/writing/agent-…

As the cost of code continues to plummet, the hard part is deciding what to build. The teams that win pick better ideas, find out sooner what customers use, and build on what works. codespeed.ai/writing/the-be…

Insane. Anthropic has replaced Git with its own version control system after Claude-driven coding pushed daily code merges from 2,500 to over 30,000 in just four months. A team of seven built Scion, a new code review tool and several related systems in five months. Scion uses 200x less CPU and memory than Anthropic's previous Git mirroring infrastructure, while CI saves 25% of CPU capacity. An early Claude Mythos Preview rewrote a hand-built pipeline to use 20x less memory and run 4x faster, with no further instructions. All the engineers who built the system now rate Claude as better at writing code than themselves.

Andrew Curran @AndrewCurran_

2 weeks ago

Anthropic's code changes merged per day has 10x since May. 'Claude is able to autonomously iterate on research loops, micro-optimize complex programs, and build new services from the ground up at a speed and scale we haven’t seen before.'

74 77 1K 132K 203

For the last 30 years, software teams planned around one constraint: writing code took time. Every roadmap, release plan, and feature cut assumed that engineering hours were the scarcest resource a company had. AI coding agents have made writing code dramatically faster and cheaper, and teams have started organizing them into software factories. In a software factory, agents take on much of the traditional software development life cycle, or SDLC, writing, testing, and reviewing code while people direct the work. Teams that build this way ship with far greater velocity, and the scarcity that once cut every roadmap down to a minimum viable product is disappearing. When writing code costs almost nothing, the real constraint becomes knowing what to build. That’s product development, and it’s now where teams win or lose with AI. The best software factories are product factories. codespeed.ai/writing/the-be…

End-to-end simulations are a new form of testing in the age of AI software engineering. They check what agent-written unit tests can't: whether a customer can actually use the product and derive the expected value. codespeed.ai/writing/end-to…

A new form of software testing has emerged in the age of AI: end-to-end simulations. Simulations automatically push an agent to zoom out because they require the agent to walk the actual customer experience end-to-end in a real environment or sandbox, revealing far more about the business value that was built rather than rote technical details that may ultimately be meaningless. End-to-end simulations check what agent-written unit tests can't: whether a customer can actually use the product and derive the expected value. An agent is given a customer to emulate and a goal, signs in to the real product, and works until the expected result is achieved. If it gets stuck, the team learns where and why. High-quality customer profiles and goal paths are especially crucial in agentic simulations. An agent that only follows the easy path, where everything goes as planned, will miss most of what goes wrong for real customers. You can also create these profiles, paths and goals with your agent team. Start with a point of view within each of your customer profiles, and probe further from there with an orchestrating agent until you're ready to run the simulations: - The analyst asking the AI assistant a question it has no data for. They ask what their biggest customer spent last quarter before the billing data has synced, and they find out whether the assistant says it doesn't know or invents a number. - The operations lead whose integration broke overnight. A connected tool sent bad data at 2 a.m., and they find out whether the product noticed, said so plainly, and let them recover. - The teammate picking up someone else's work. They open a half-finished workflow with no context, and they find out whether the product explains what happened and what comes next. - The power user on their fortieth task of the day. They use keyboard shortcuts, bulk actions, and three tabs at once, and they find every place the screen falls out of sync with what actually happened. - The new admin on day one. They connect the company's tools, invite the team, and import last year's data, and they find every step that assumes something already exists. - The skeptic. Every time the product says saved, synced, or sent, they go and check that it actually happened. Read the complete post: codespeed.ai/writing/end-to…

Unit tests used to be the gold standard in software engineering. Great unit tests captured both a product and business perspective in addition to the technical requirements needed for successful functionality. But agents can't bring that same perspective continuously, and even the best frontier models as of 2026 produce far too many unit tests with no real purpose or value, creating immense bloat and slowing both agent and human teams down. For example, an agent will write ten unit tests proving a string contains its own substrings. So to the agent, all ten pass, but nothing actually proves the product works. Part of the issue is that agents sometimes lack the perspective to consider their work more broadly, and are therefore inclined to write positive tests they know will pass instead, because the only context they have is the code and discussion in front of them. That makes these unit tests worse than harmless: they look like proof that the product works when nothing has been checked. To avoid the pitfalls that have emerged in testing agentic code, a new form of software testing has emerged in the age of AI: end-to-end simulations. Simulations automatically push an agent to zoom out because they require the agent to walk the actual customer experience end-to-end in a real environment or sandbox, revealing far more about the business value that was built rather than rote technical details that may ultimately be meaningless. Read the full post: codespeed.ai/writing/end-to…

The next frontier in AI software engineering lies above and around the models. Teams are evolving into roles that combine product, engineering and vision. At @codespeedai we’re building intelligence systems, autonomy and governance, team tooling, and more. And the only way to build at an incredibly fast pace is through deep product intuition and engineering skill.

Harness & intelligence engineering at scale is incredibly tedious but also deeply invigorating. Some of the most exciting work pushes far beyond deterministic tech and into all of the probabilistic ways humans and agents think into a collective mind that operates at scale.

While there are many similarities between AX and UX, agents consume data and functionality exponentially faster than people, around the clock. As a result agent experience optimization tends to focus on API design, tools and abilities, autonomy, and governance. At Codespeed we're particularly focused on filling the gap in multiplayer functionality and interfaces for people and agents. Most agent memory is single-player, held by one agent and the person who briefed it. Multiplayer changes that. The whole team and the whole agent workforce collaborate in the same threads, and Codespeed Intelligence forms the collective memory that carries those decisions into future work. Agents then begin to form understanding and opinions about the business itself, allowing them to proactively surface potential improvements not only to a codebase, but to the broader business, without ever being asked. codespeed.ai/writing/ax-is-…

Most agent memory is single-player, held by one agent and the person who briefed it. Codespeed multiplayer changes that. codespeed.ai/writing/ax-is-…

Codespeed's core intelligence engine revolves around transmuting human taste and judgement into agentic infrastructure as a permanent part of the machine that powers autonomous workforces. Not a step from which to eventually graduate. Even if agents are able to achieve a high degree of autonomy, taste and judgement will always be core to the individuation and thus value of any creation. x.com/fchollet/statu…

François Chollet @fchollet

2 months ago

I believe we need to make a deliberate effort to keep humans in the loop in all critical processes across our economy and society, regardless of whether it is technically necessary. Even if AI develops the *capability* for advanced autonomy, we should not make it highly

202 225 2K 109K 337

The next frontier in agentic intelligence is about work origination. What should be done & why. Traditional hierarchical constraints don’t exist when you can spin up incredibly intelligent teams that instantly contain the collective knowledge & memory of all people and agents in your company, including discernment to know what’s valuable & what isn’t.

gabriel @gabriel1

2 months ago

i always thought of work as “work exists and someone is hired to do it” but work doesn’t just exist, someone has to come up with it, and that’s where nearly all economic value is created that’s why top-down hierarchies lose, a million decisions can’t all be made from the top

12 20 552 29K 142

There is a huge delta between tight startup teams who ship mercilessly fast and the technical leads that help get them there, and "engineering management" as we know it in big tech. The former is where agentic team development is heading. x.com/GergelyOrosz/s…

Gergely Orosz @GergelyOrosz

2 months ago

As someone who had been an engineering manager in the past, it's laughable for anyone to compare AI agents management to people management. Engineering management is as close to this as The Minions and Gru is to the real word (and AI agents *are* like those Minions, seriously)

52 17 618 82K 107

Graphical user interfaces are the crucial evolutionary next step for agentic coding. They unearth an incredible amount of missing value in current AI workflows for software engineering, from governance to quality to intelligent feedback systems. What we’re building at Codespeed is truly the first of its kind.

Patrick Collison @patrickc

2 months ago

I love agentic coding harnesses, but they shouldn't be primarily terminal-based. The terminal is great for quick and precise commands, but information density is extremely low and UI affordances are minimal. Maybe provision of TUIs is worthwhile for occasional use (when

539 285 6K 883K 2K

As software engineering evolves, it's clear that writing code is far less of the essential equation. Instead effective communication, management and strong leadership equally in business and tech are replacing it. The engineers who bring a clear vision, communicate precisely, and manage their agentic teams in the same ways the highest output human teams have operated at top companies for decades will continue to see incredible compounding results. In 2026 and beyond, AI software development is far more than a set of tools and features. It's a force multiplier for the best leaders on your team.

The age of agentic software engineering has spanned just eight months, starting with the advent of the frontier models that changed the industry in 2025. Yet the acceleration during this time has created a period of compression as if years have already passed. As a result "classic" problems in agentic software development have emerged as millions of engineers experiment with workflows, systems, tools and models every day. The largest of these problems is a very human one: one person has only so many hours in the day, while a team of agents can be spun up to multiply work effort instantly. As the rate of software development output has increased and the demand for great software engineers who can communicate and work with agents continues to grow thanks to Jevons Paradox, leading to more not fewer jobs, the need to reduce human friction and bottlenecks has become a forefront problem to solve. What's emerged are often manual workflows that include careful skills handed to an orchestrating agent equipped with the best of frontier models who carefully manage sub-agents beneath them, checking their work, output quality, and holding the line for security. But that process still requires an incredible amount of maintenance, it isn't set up for multiplayer, agents don't learn across a company, and as a result a brittle software factory system forms that can't easily adapt to the near lightning speed of AI software development today. Codespeed changes that. Every project gets a Lead Dev: one agentic manager that runs the whole effort, working at every altitude from scoping to review, and carrying into each moment only the context that moment needs. The bottleneck, in other words, is a choice now. And software is not the first industry to reach this point. The car industry got there first.

The reason GUIs and great UX will continue to dominate in an agent-first world is that they’re fantastic tools for capturing the infinite touch points that comprise human taste and judgement. Which it turns out is the precise probabilistic and deterministic infrastructure agents need to do great work as they become increasingly autonomous. Every team’s definition of what good means is vastly different. It’s a cascade of small decisions, judgement calls and discussions that go from rough-around-the-edges to a clear vision over time. A graph of actions that is impossible to capture without the context and web of thinking around it. Capturing judgement and taste as infrastructure is exactly what we’re building at @codespeedai, and it’s how we’re powering the next generation of truly autonomous agents.

Data centers have been part of American life for thirty years. And all future benefit to humanity by AI depends on them. x.com/i/article/2086…

AI software development has solved capacity, not leadership. Codespeed Lead Dev scopes the work, holds all agentic changes to the standard your team decides, and brings you only the calls that need you. codespeed.ai/writing/the-se…