Software that gets easier to change as it ages
Every codebase gets harder to work in over time, because the reasoning behind it decays faster than the code does. Keep the reasoning and that stops being inevitable.
Compound Engineering
Recursive self-improvement does not arrive. It accumulates.
Recursive self-improvement is not waiting on a bigger model. It is waiting on work that accumulates, where every decision, correction, and standard becomes something the next run starts from rather than something the next run rediscovers. Engineering has never worked that way. Compound Engineering is the argument that it can, and that this is where takeoff actually begins.
Where This Goes
Every law you know about classic software engineering assumes nothing accumulates (except tech debt).
An engineering organisation that gets faster as it grows
Adding people slows a team down because each new person has to rediscover what everyone else already knows. That cost only exists while the knowing lives in heads.
Judgment that outlives everyone who made it
The taste of the best engineer you ever hired currently leaves with them. It does not have to. What they held the bar to can still be holding it in ten years.
The last time your team solves this problem
Most engineering effort goes into questions the company has already answered somewhere. Answer them once and the second occurrence stops being work at all.
What Changed
AI made your team faster at forgetting.
An agent writes more in an afternoon than a team used to write in a week, and starts tomorrow knowing none of it. Throughput went up and retention did not move, which means the same mistakes arriving faster and a review queue nobody can hold. More output from a system that forgets is not leverage.
Recursive Self-Improvement
Your AI workforce gets better every night.
Codespeed Dreaming gives your AI workforce what every intelligent system needs: time to consolidate what happened, decide what matters, and wake up better.
The human mind does some of its best work while we sleep. Memory settles, noise falls away, patterns surface, and the mind returns sharper.
Codespeed brings that rhythm into autonomous AI software engineering teams. After your specialists and Lead Dev build, review, investigate, and learn, the workforce sleeps on the work.
The result is a workforce with an inner life. One that learns what to remember, what to forget, and how to think with more clarity and focus for the next day.
A memory that doesn’t change the next run is just stored context. Dreaming turns it into judgment.
Memory
Reviewed every night, under the rules you set.
Thread
Acceptance
Correction
Decision
Kept
Let go
Your controls
Sharper work
01
Keep what earned trust
The lessons your team stood behind become how specialists work tomorrow.
02
Let the noise fall away
What the work stops confirming loses its hold, so yesterday’s guesses don’t steer tomorrow’s decisions.
03
Wake up sharper
Specialists and Lead Dev return with clearer context, calmer judgment, and a firmer sense of your quality bar.
Dreaming runs on its own by default
Left alone, Dreaming runs every night and cleans up after itself, settling what your specialists remember and dropping the assumptions Lead Dev should stop carrying forward. If you want a closer hold, you can require review before any memory changes, or turn it off completely, either for the whole project or for one specialist.
Memory follows the same permissions as the work
Memory runs on the permissions you already set, so governance always runs end-to-end. Dreaming never gets more freedom than the specialist already has.
Every memory stays visible and correctable
Every memory shows what the specialist learned and how much weight it carries. Remove anything wrong or out of date and the next night takes the correction into account. Leave it alone and Dreaming still prunes what stopped earning its place, so you can stay as hands-off or as hands-on as you like.
The Mechanism
Your lifecycle is the training environment.
Not a benchmark and not a fine-tune. The change your team accepted, the correction they wrote, the review they sent back, and the standard they held it to are all signals about what good means at your company. A system that captures them gets better with use. A system that does not is identical on day four hundred and day one.
What Cannot Be Bought
Everyone gets the same models. Nobody else gets your judgment.
The labs sell the same weights to your competitors, so whatever advantage lives inside the model is an advantage everyone has by Thursday. What your team decides, corrects, and holds the bar to is the part nobody can buy, and it lives in your own repository in an open format. It keeps accumulating whatever you choose to run it with.