Retrieval answers a question. The next challenge is making that answer part of a repeatable operation: something that can run when new information arrives, produce a predictable result, and deliver it to the team that needs to act.
Our latest release adds that operational layer to Knowledge². Agents, feeds, subscriptions, pipelines, destinations, and typed Envelopes now compose directly on top of governed collections. Work that previously crossed several integrations can be described, tested, and operated as one connected workflow.
This article is adapted from the original Knowledge² post on LinkedIn.
From a retrieved answer to a continuous operation
Consider support-ticket triage. Historical tickets arrive in a collection on a schedule. As each new ticket lands, a pipeline extracts severity, product area, and customer tier into a structured Envelope. A feed then evaluates those Envelopes against routing rules: an outage affecting a high-severity service can move to Slack and Jira, while an issue for a top-tier account can escalate to PagerDuty.
The routing itself is straightforward. The important change is architectural. The collection, extraction step, event stream, and delivery rules share one model rather than being distributed across a help-desk plugin, ETL job, vector store, workflow engine, and the glue between them.
The same primitives support work where reading matters more than routing. A market-monitoring collection can follow competitor release notes, pricing pages, and product updates. When a source changes, a query agent compares the new material with your positioning, identifies what matters, and keeps each conclusion tied to its evidence. Relevant changes can go to Slack or a weekly digest; everything else remains searchable.
These look like different workflows, but their shape is the same: knowledge arrives, an agent performs bounded work, and a typed result moves to the right destination.
Consolidation without black-box autonomy
This is not a generic automation canvas with an LLM added at the end. Each workflow runs against governed collections, produces typed outputs, and retains provenance as information moves from ingestion to action. When a request spans several sources, the system can route by intent, call the relevant agents, and combine their results with citations.
That does not mean handing judgment to the model. The system handles repetitive work around knowledge—extracting fields, classifying items, summarizing changes, and routing results—while people remain responsible for decisions that require context, authority, or consequence.
Six primitives, one shared contract
- Agents perform bounded work. A typed agent can query, extract, classify, summarize, review, or notify against a governed collection.
- Feeds make the work continuous. They run an agent on a schedule or react when new information lands, turning an on-demand capability into an ongoing process.
- Subscriptions decide what matters. Structured filters and semantic matching determine which events should move forward and where they belong.
- Pipelines describe the topology. Collections, agents, feeds, and their connections live in a declarative specification that can be dry-run before it is applied.
- Destinations deliver into existing tools. Results can reach systems such as Slack, Jira, ServiceNow, or PagerDuty, with signed webhooks available for everything else.
- Envelopes keep results typed and traceable. Content, extracted fields, metadata, and provenance travel together in a consistent shape.
The shared Envelope contract is what makes the pieces composable. Each primitive understands the same payload, so teams do not have to translate data at every boundary. Low-code here is not a drag-and-drop abstraction; it is a small set of components designed to agree with one another.
Retrieval remains the foundation
None of this replaces retrieval. It depends on retrieval being dependable.
Hybrid search, cross-encoder reranking, per-corpus tuning, and passage-level citations remain underneath every workflow. If the system cannot find the right source material, a polished orchestration layer only moves the wrong result faster. That is why Knowledge² shipped the retrieval foundation before adding the workflow layer above it.
A good retrieved answer resolves a question at a point in time. Composable workflows extend that capability across time: they watch for new information, apply a defined operation, preserve evidence, and deliver a result where work already happens.
Start with one useful loop
The best first workflow is deliberately small. Choose one governed collection, give one agent a bounded task, connect one feed, and deliver the result to one destination. Once that loop is observable and useful, add the next source, rule, or subscriber.
The goal is not to model an entire organization on day one. It is to remove one repetitive knowledge bottleneck with a system your team can understand, inspect, and extend.