Software architecture platform · Prelaunch
Continuum keeps your system’s boundaries, business rules and decisions in one model that people and AI coding agents can read and check before they change code. Keep your architecture decisions with the project.
What goes wrong today: the next engineer or agent session sees the code, not the reasons. In HashRelay, the architecture said missed notifications were kept until acknowledged. The code expired them after 24 hours, and the dashboard still counted them as delivered.
For architects, developers and engineering leaders whose teams write code with AI agents.
A delivery queued for an offline device is kept until it is acknowledged or its replay window (24 hours by default) elapses. An expired delivery is never counted as delivered.
The decision this diagram explains. It was proposed by an AI agent and accepted by the project owner, with the alternatives and risks on record.
The platform
Continuum is a software architecture platform. Its core is an explicit model of how your system works, written in two small languages. Decisions, checks, diagrams and revision history all point back to the same model, so a person or an agent who opens it next week starts from the same place you left.
Plain, versioned source that lives with your project.
Decisions
Each decision records its rationale, the alternatives considered and the risks, plus the model elements it expects to exist.
HashRelay: ADR-001 names the requirement and the service that must satisfy it.
Validation
The compiler resolves every reference and runs its rule checks. It also flags intent nothing realizes yet.
HashRelay: 78 checks passed, 0 errors, 18 review warnings.
Diagrams
Diagrams are drawn from the compiled model and checked against it, with inferred parts marked.
HashRelay: the replay diagram above, 0 claims drawn stronger than the model supports.
Revision history
Changes are written as change files against a pinned model state, so the next session knows exactly what it builds on.
HashRelay: a proposal six days later started from the state ADR-001 produced.
Why it matters
AI agents write code quickly, then forget. A new engineer reads the code and has to guess. Either one can “fix” behaviour that was a deliberate decision, or keep a promise the system never actually made. A shared model gives everyone the same answer to what did we decide, and why?
Decisions stay attached to the elements they govern, with alternatives and risks, instead of fading in a document nobody opens.
Read the accepted requirement before changing code, and start the next session from a pinned model state rather than a summary.
See what is decided, what has been checked and what is still unknown, each labelled for what it actually establishes.
Proof · HashRelay
HashRelay sends notifications from scripts, CI jobs and AI agents to devices subscribed to a hashtag. Avlon built it with Continuum. The case study follows one real decision from the mismatch an agent found to the tests on either side of the 24-hour boundary.
Change review
The decision check found its two expectations missing from the model. After the change, both held.
Either side of 24 h
Regression tests pin the boundary. The test that expired deliveries are never counted as delivered failed before the fix.
Next session
Six days later, an agent drafting macOS and Linux support started from the same model state and kept the replay rule as stated.
Status
Continuum has not launched yet. Here is where each part stands, based on Avlon’s own project records.
| Capability | Status | Evidence |
|---|---|---|
| ADL and BDL models, compiled and validated | Used on Avlon projects | HashRelay’s model compiles with 78 checks passed and 0 errors. |
| Signed decision records with model checks | Used on Avlon projects | HashRelay’s record holds 19 signed events and 3 decisions. |
| Diagrams drawn and checked from a model | Used on Avlon projects | The replay diagram on this page was generated and verified for this site. |
| One project shared across ChatGPT and Claude Code | Development deployment | Round trip demonstrated on 15 Nov 2026 with a small test project. Not released. |
| Proof that code conforms to the model | Not claimed | Continuum records implementation evidence as reported. It does not certify code. |
Walk through one decision from question to code, then ask to hear when you can try Continuum on your own project.