SynthOrg Documentation¶
Describe a piece of software, and the work is split into parts that stand alone, built in parallel, and each part checked by something that did not write it. On your hardware, against the models you choose.
Pre-alpha, and not yet fully working
The loop has been driven live twelve times and has never reached the assembly stage: no run has produced an assembled deliverable. These pages describe how the system is built and how it behaves, which you can check against the code. Expect bugs, missing polish, and breaking changes. See the Roadmap for what is wired versus what is intent.
The problem¶
An agent working alone cannot hold a whole application: it does one thing at a time, and the twentieth damages the first. Adding agents does not fix that, because the binding constraint is not agent supply but decomposition quality. Splitting work so the parts are genuinely independent is the hard problem, and the one this system is built around.
The mechanism¶
A tree does not have the single-agent failure mode, provided the merges hold. So an objective is decomposed into units that can each be built alone, the leaves build concurrently, each in its own workspace, and the tree is assembled from the bottom up.
- Decomposition turns an objective into a tree of units, each with declared dependencies and expected artefacts. A unit that is not atomic is split again. See Recursive Decomposition.
- Dispatch builds waves from the whole tree's dependency graph, so independent subtrees run together. A unit whose declared inputs died is parked with the reason, never dispatched onto dead work. See Coordination.
- Execution runs each unit in its own scoped workspace, sandboxed from the host, with a governed tool surface. See Agent Execution.
- Review is done by something that did not write the work: the executor is excluded from selection, and the archive refuses a verdict whose reviewer and executor match. An independent check is a triage filter, not an authority. See Verification Quality.
- Assembly is a gated stage of its own, not a side effect of the last unit finishing: an accountable assembly task, then an evaluation pass scored against the plan's own criteria. See Initiative Tail. That is the loop as built; The Build Loop is the authority on where it goes next, and retires assembly as a stage in favour of merging each unit onto a trunk that stays green.
Where it runs¶
Self-hosted, on your hardware. Every LLM dispatch names its own explicit (provider, model)
pair, including local models, with no default to fall back on. No SynthOrg service sits in
the path, so on the model path what leaves the machine is what your chosen provider needs,
and against local models nothing does. Two outbound integrations exist and both ship off:
product telemetry and
cross-deployment analytics, each sending only what its own
page documents, and only once you enable it.
Get Started¶
-
Run it
Stand up the platform with Docker and configure it through the dashboard.
-
Quickstart
Install the CLI, start the stack, and complete the setup wizard.
-
Develop it
Clone the repository, set up a development environment, and contribute.
-
Guides
Configuration, providers, budgets, security, deployment, and operations.
Design Specification¶
The design pages are the source of truth for designed behaviour. Each area marks its own wiring status; treat any gap between a page and the code as the work, not the spec.
-
Design Overview
What the system is for, its principles, and the vocabulary the other pages use.
-
Recursive Decomposition
How an objective becomes a tree, the atomicity gate, and the recursion bounds.
-
Coordination
Dependency-gated waves, parallel dispatch, parking, and run recovery.
-
Verification Quality
Who reviews, why they cannot be the author, and what a verdict does and does not mean.
-
Initiative Tail
Assembly and evaluation: the only path by which a plan reaches completion.
-
Agent Execution
The execution loop, sandboxed tool use, context budget, and termination.
-
Providers & Budget
Explicit provider binding, cost recording, and spending controls.
-
Security
Autonomy levels, the approval gate, sandboxing, and untrusted-content fencing.
What is wired¶
Every item below is implemented and exercised by deterministic end-to-end harnesses driven by a scripted provider, so the suite involves no real spend. That is a statement about the code paths, not a claim that a run delivers.
- Decomposition and dispatch: recursive planning, a dependency graph over the whole tree, parallel waves, and level-triggered recovery after a restart.
- Independent review: roster-staffed gate roles, executor exclusion, and archived verdicts recording the model each judgement was made under.
- Sandboxed execution: each unit in its configured sandbox, a scoped shared workspace, and a governed tool surface including an MCP bridge.
- Explicit provider binding: an unconfigured feature is off and says so rather than borrowing a default.
- Budgets and cost attribution: per-agent limits, run and token ceilings, spend attributed by prompt purpose.
- Approval gate and audit: actions parked for a human by autonomy level, with a signed audit chain.
- Memory and knowledge: per-agent and shared memory with a retrieval pipeline, plus an ingested corpus agents can cite.
- Operator surfaces: a React dashboard, a REST and WebSocket API, and runtime-editable settings.
Roles, departments and staffing are shipped machinery. They are plumbing: how a reviewer is found and how permissions are scoped, not the reason the system works.
Further Reading¶
| Section | Description |
|---|---|
| Architecture | System overview, module map, design principles |
| Tech Stack | Technology choices and engineering conventions |
| Decision Log | All design decisions, organised by domain |
| REST API | REST + WebSocket API reference (Scalar/OpenAPI) |
| Library Reference | Auto-generated from docstrings |
| Roadmap | Status, open questions, future vision |
Links¶
- GitHub Repository
- License (BUSL 1.1, source available; free production use for non-competing small organisations; converts to Apache 2.0 at the Change Date)
- Licensing & Usage: what is permitted, why BUSL, and how to get a commercial license