Research70% lossless compression2× faster inferenceContinual learning research

Software that improves.
Intelligence that learns.

Build, ship, learn from real usage, and improve. All in one connected platform.

Get startedExplore research
feat/retry-budget
CodeGraphAtlasTasksJobs
MRJO2
Tools

Explorer

services
checkout
BUILD.rbs
client.go
client_test.go
retry.go
payments
web
.reasonos
README.md
WORKSPACE.rbs
retry.go
client.go
1package checkout
2
3import (
4 "math/rand"
5 "time"
6)
7
8// Policy is how the checkout client retries calls
9// to the payments gateway.
10type Policy struct {
11 Base time.Duration
12 Ceiling time.Duration // never wait longer than this
13 Max int
14 Rand *rand.Rand
15}
16
17// Backoff is the wait before retry number attempt.
18func (p Policy) Backoff(attempt int) time.Duration {
19 d := p.Base << attempt
20 if d > p.Ceiling || d <= 0 {
21 d = p.Ceiling
22 }
23 return d
24}
25
26// Retry reports whether a call that failed with
27// status should be tried again.
28func (p Policy) Retry(status, attempt int) bool {
29 return attempt < p.Max && status >= 500
30}
AI
Chat 1

What do you want to do today?

Pick one, type its number, or describe what you need.

  1. 1Build or change somethingDesign before building
  2. 2Find and fix a problemDetect and fix issues
  3. 3Review my workChoose the project, period and evidence
  4. 4Teach meExplore RBS capabilities through guided demos
  5. 5Explore this codebaseUnderstand the architecture
Message ReasonOS
Ask: commandsClaude Opus 5
RBS
Ready
UTF-8
GO
✦ Completions
◆ Hide AI
◀ Hide Explorer
⫿ Split

The problem

Your tools don’t share what they know.

The error is in one tool, the trace in another, the code in a third and the customer impact in a fourth. Every person, and every agent, rebuilds that picture by hand before anything gets fixed. Most of what a release could teach you is lost between them.

The loop

Every release teaches the next one.

ResearchModels that learn from the loop itself
  1. 01BuildYour team and coding agents
  2. 02TestQA agents on web and mobile
  3. 03ShipCI builds what changed
  4. 04ObserveErrors, traces, infrastructure
  5. 05UnderstandWhat customers actually did
  6. 06LearnYour team decides what to change
  7. 07ImproveAgents make it, people approve it
  • In the product today
  • People decide, agents assist
  • Research

See it work

One turn of the loop.

Checkout errors climb. The trace shows the payment gateway failing and being retried.

Any serviceLast 1 hour
Duration · last hour · 797 traces
TimeStatusOperationDurationSpansServicesErrors
14:02:11.482OKcheckoutGET /cart38 ms62—
14:02:10.907ErrorcheckoutPOST /charge412 ms1443
14:02:10.311OKinventoryGET /stock/:sku21 ms31—
14:02:09.854OKcheckoutGET /orders/:id64 ms83—
14:02:09.120OKpaymentsPOST /refund133 ms93—
14:02:08.733OKcheckoutPOST /charge188 ms114—
14:02:08.045OKledgerPOST /append47 ms51—
14:02:07.598OKcheckoutGET /cart41 ms62—
14:02:06.990OKcheckoutPOST /charge205 ms114—
14:02:06.211OKinventoryPOST /reserve29 ms41—
14:02:05.876OKcheckoutGET /health2 ms11—

Use it today

Start with what you need.

All products
Also availableOrchestrationEditorHosted GitStacked changesAtlasBuildRemote cacheRemote executionInfrastructurePlanningDocumentsCollaboration

Research

AI that gets better with experience.

Today’s AI is a few large, general models that are expensive to run and the same for everyone. We are researching efficient, specialized models that learn from the work they do, and could become intelligence that belongs to a person or an organization.

Explore our research
Model compression · internal result70% lossless compression2× faster inferenceWithout quantization.
Continual learning · working demoModels that improve through experience.Our General Learning Operator updates a model’s weights from completed work. In internal tests it got better at later coding tasks.

Why trust us

Real software, shipping every week.

We built the stack underneath

Our own build system, remote cache, hosted Git and a server for every branch. When something is slow, we can fix it at the bottom.

How builds work

ReasonOS ships ReasonOS

Our own code is built, tested, reviewed and released on the platform, with our own agents working in it.

Read the changelog

Research with a technical record

Each result lists its objective, method, baseline, limitations and how to reproduce it.

Inspect the results

Latest releases

All releases

  1. Production

    Infrastructure monitoring

    See hosts, clusters, nodes, workloads, pods and volumes next to your traces and logs, with built-in alerts such as a disk filling up.

  2. Production

    Agent traces

    Follow each model call, tool call and its cost in a coding agent’s run, in the AI section of Observe.

  3. Production

    Observe and Understand

    Logs, metrics, traces and errors in Observe. Product analytics, session replay, feature flags and experiments in Understand.

Enterprise

AI agents under your control.

Security and administration

Single sign-on

SAML or OIDC through your identity provider. Two-factor is on by default.

Custom roles

Allow and deny rules for people, teams and service accounts.

Audit log

Every sign-in, push, merge, permission and secret change, with who did it.

Agents under your rules

Agents get the same permissions, required checks and reviews as people.

Start with one product. Grow into the platform.

Get startedTalk to us