Each product stands on its own. These are the workflows that only work because they share one graph — where memory, code context, routing and validation are the same state, not four systems agreeing with each other.
A fresh agent session begins blind: no memory of last week, no idea why the retry loop is odd. Wire CTXone and AgentStateDeveloper into the same MCP connection and the session opens with the team’s decisions and the code’s own ledger already in hand.
$ ctx recall "payments architecture" $ asd ledger get payments.chargeCard
Multi-week work dies at the session boundary — the plan drifts, the reasoning evaporates, and the next session re-derives decisions that were already settled. A plan held in the graph is state: resumable, checkable, and branchable when you want to try a direction without losing the current one.
$ ctx plan next $ ctx branch spike/new-schema
In regulated environments the question is never just "did it work" — it is who authorised it, under what policy, and can you prove the record was not edited afterwards. Provenance on every write, policy gates on sensitive paths, and a hash-chained audit log make that answerable rather than aspirational.
$ ctx why-did-we /memory/legal/retention $ asd audit verify
Public benchmarks rank models on someone else’s tasks. Branch the same conversation across several providers to compare them side by side on work you actually do, then let routing history turn that comparison into a standing decision per task type.
# fork one thread across providers, # then route on the observed result
You changed a prompt and one run passed. That is a single sample from a distribution you have not examined. Run the scenario against multiple agents, capture each decision path as a graph, and check the blast radius of the underlying code change in the same pass.
$ asd effects_of processOrder # then replay the scenario and diff runs
Blanket downgrades to a cheaper model save tokens and cost more overall, because the retries land somewhere you are not measuring. Route on cost-per-unit-of-work drawn from your own history, with budget circuit breakers, and keep every routing decision as a commit you can interrogate later.
# route by task type on observed # completion cost, not token price
Nothing here requires the whole suite. Each product is self-hosted and MCP-native on its own — they compound when they share a graph.