Aggregate
Substrate to packet
Pulls the structured data each stage needs from Isilon, validates it, hashes it. The packet is the contract between stages.
§ 20Stack · Atelier
Atelier. The orchestration layer the rest of the Stack runs on.
Multi-stage AI as a first-class primitive. Aggregate the substrate. Hand a clean evidence packet to each stage. Hash it. Cache it. Write a ledger row. Make every claim traceable. This is not a sibling product — it is the harness the rest of the Stack quietly runs on.
§ 01 / The problem we are solving
One AI call works. The fifth one in a row breaks everything.
The first AI call in a workflow is easy. The fifth one — the one that depends on the third one's output, which depended on a join the second one made against a cache that the first one populated — is where most teams quietly fail. Output drifts between calls. Cache invalidation gets accidental. Source attribution disappears two stages in. Three months later nobody can reconstruct why the agent said what it said.
Atelier is the layer we built to stop that from happening. Every multi-stage AI workflow we run goes through it. The stages can swap; the harness does not.
§ 02 / What it does
Four jobs. One harness.
Orchestrate
Stage graph
A typed pipeline of AI stages. Each stage names its inputs, outputs, and the model class it expects. Terra grounds, Soqratic reviews, custom stages compose. Failures are local; retries are typed.
Cache
dataHash-keyed
Same evidence in, same output out. The packet hash gates re-runs. A small change to the substrate invalidates only the stages that depended on the changed slice.
Ledger
What ran, what saw, what said
Every stage writes a row: input hash, output hash, model, tokens, cost, citations. Six months later the question “why did the workflow produce that” takes one query, not one engineer.
§ 03 / The pattern, in practice
One workflow, end to end.
A facilitator on our program clicks “synthesize this session.” What runs:
- Isilon assembles the evidence packet — transcript, prior sessions, journal entries from the relevant week, the student's active program. The packet is hashed.
- Atelier reads the hash, checks the ledger. If the same packet has been processed by the same pipeline already, we serve from cache and write a hit. If not, the stage graph runs.
- Terra grounds claims against the packet. Each claim gets a source-attributed citation or a contested flag.
- Soqratic reviews the grounded draft, asks the next right question, surfaces what the draft is avoiding.
- A final synthesis stage emits the artefact, with citations intact, into the facilitator's
artefacts/folder. - The ledger writes one row per stage. Observatory indexes it.
The facilitator sees a synthesis appear. The audit trail is automatic. If the transcript later changes, only the stages that depended on the transcript re-run. If the synthesis prompt changes, only the synthesis stage re-runs. Nothing is recomputed by hand.
§ 04 / Where it sits
Underneath. Always underneath.
Atelier is not a SKU you would buy alone. It is the harness an Avira install deploys into your runtime so the rest of the Stack can do its job consistently. Isilon hands it packets. Terra and Soqratic are stages it orchestrates. Observatory indexes its ledger. Custom stages — the ones we write for your domain in week 2 of an install — slot in as first-class.
We are listing it on the Stack rather than burying it because the orchestration pattern is the differentiator. The reason our program can run on this software is Atelier. The reason your team will be able to is the same.
§ 05 / What it is not
Three things to be clear about.
Not a no-code workflow builder. Stages are written in code, by engineers who know the domain. The harness runs them. If you wanted Zapier, this is not Zapier.
Not a generic agent runtime. Atelier orchestrates AI stages — typed, evidence-bound, citation-emitting. It is not a tool-calling loop with autonomy. Different shape, different problem.
Not yet a public SDK. Atelier ships inside an Avira install today. The library extraction is in progress; if you want to be the first install asking for it, write to us.
Stop running multi-stage AI without a harness.
If your workflow has more than two LLM calls and nobody has written down how the cache invalidates, this is the layer you are missing.