The software development platform
Skynet would have killed for.
Stable, scalable, and hands-free tools for building the next generation of software, intended to replace an obsolete stack not designed for the new world.
State of the world
As more and more teams move to fully agentic work...
-
Pull requests and code review have become pointless
Agents now write far more code than people can read. Nobody can realistically review all of it, so a review becomes a stamp: LGTM.
And review never really helped find bugs in the first place. In Code Reviews Do Not Find Bugs, researchers at Microsoft report that reviews often miss the functionality issues that should block a change.
-
CI and merge times are skyrocketing
Git and GitHub were never designed for massively parallel work. Now even a very small team can easily run hundreds of agents at once.
Their changes wait in one merge queue. Each change that lands makes the others integrate and run CI again, so the more agents you add, the longer each change waits.
-
Stars and social proof have lost their meaning
Social coding taught us to judge code by its stars, forks, and followers. Now code is cheap to produce in very large quantities.
Stars measure popularity. They don’t tell you whether the code meets your needs.
Amp
Version control for massively parallel agentic work
Git represents versions as snapshots linked by commit ancestry. Amp represents versions as sets of patches with explicit dependencies.
-
Order-independent integration
The same set of patches produces the same code and conflicts, regardless of arrival order. Agents work concurrently without you managing integration order.
-
Resolve once and share it
A resolution is another patch. Agents exchange it with the conflicting changes, so each channel doesn’t need to solve the same conflict again.
-
Import from Git
Bring your existing repository into Amp, including branches, tags, and history. Keep the context behind your code as your agents move to a workflow built for parallel work.
Gauss
Repositories built for fully automated development cycles
GitHub centers code hosting on human participation. Gauss centers it on agents exchanging code under explicit channel policies.
-
Agents self-manage
Agents can submit changes, inspect failures, and push fixes through commands and APIs. Nobody needs to navigate a website or shepherd a merge.
-
Automatic checks and merges
All changes go through a merge queue designed to leverage Amp’s unique commutative properties, ensuring build retries are kept to an optimal minimum when automatic checks are in place.
-
No social proof
Stars measure attention. Approval tells you little when the human hasn’t inspected the code. Gauss bases integration on explicit channel policies and executable checks.
Piston
Remote compute for every agentic need
Generating code uses inference; building, testing, and running it needs compute too. Piston takes that work off your laptop. Agents send it the files they’re editing right now, or run a workflow from your repository.
-
Run changes without pushing
Agents can submit their current files, including unrecorded edits, instead of invoking a workflow against a repository ref. Of course, Piston tasks are designed to be used for CI as well, running as Gauss checks against merge candidates.
-
Fully isolated and fully informed
Piston runs builds, tests, scripts, and other tasks in isolated environments. Agents can easily retrieve structured logs, results, and output files to decide what to do next.
-
Agent work runs remotely
Fast remote execution keeps their work moving without using your laptop’s CPU and memory.