The Agent Stack.
The infrastructure an autonomous financial agent actually runs on — observability, orchestration, retrieval, lineage, constraints. Every entry is described for what it contributes to governing an agent that touches capital, not for what it does in general. That framing is the whole register: there is no shortage of writing about these tools, and almost none of it asks the question a risk officer asks.
How to read this register.
Entries are published at one of two levels, and the level is stated on every page:
- Profile — scored against five published criteria, with a defended caveat. Requires that someone has actually used it in anger.
- Listed — not yet reviewed — structured facts and a governance lens, with no score. Honest about what it is.
The second category exists so the register can be complete without pretending to judgement it has not earned. If you build one of these and want a full profile, walk us through it.
No entry in this register has a commercial relationship with the QIS Ecosystem. Listing is editorial and is never sold. Where that changes, it will be disclosed on the entry itself and in partners.json.
Analytics engine
Constraints
Programmable validation and correction applied to model inputs and outputs.
Data quality
Declarative data validation — expectations as executable, versioned assertions.
Evaluation
Experiment tracking
Graph database
Ingestion
Lineage
An open standard for collecting dataset and job lineage metadata across tools.
Observability
Traces, prompts, and evaluations for LLM applications — the closest thing to a decision trail an agent currently produces.
Production ML observability with drift and performance monitoring across model versions.
Orchestration
Microsoft's SDK for embedding model orchestration into conventional applications.
Graph-structured agent workflows with explicit state and controllable execution paths.
Programmatic optimisation of language model pipelines — compiled prompts rather than hand-written ones.
Durable execution — workflows that survive process failure, with complete replayable event history.
Asset-oriented data orchestration with typed dependencies and materialisation history.
Research environment
Reactive Python notebooks stored as plain files, with deterministic execution order.