Jylus @jylusdev

@hwchase17 I'd give background agents their own identity and explicit delegated scope. But the audit also needs the evidence available when they acted. That's the part Jylus supplies: what was known then, not just what's in the database now.

@levie I'd put evidence tokens per completed task on the swarm dashboard. In our frozen 528-question test, Jylus averaged 2,532 input tokens vs 205,129 for full context with the same model. How much of the swarm's spend is rereading overlapping evidence?

@nicbstme The feed needs to retire its own suggestions. A renewal draft becomes wrong when the customer cancels an hour later. How do you decide which cards to refresh? Jylus handles the state/evidence side; the app still has to update the card.

How much of your agent bill is spent reading irrelevant data? Our frozen 528-question test, same model. Average input tokens/query: Full context: 205,129 Jylus: 2,532 Strict accuracy: 78.79% → 100% on this test. jylus.ai/benchmark-meth…

@LangChain @snyksec The escalation path is the bit I'd inspect. "No matching source" and "the search didn't finish" need different outcomes. Jylus exposes search and evidence completeness separately. How does Assist route those cases?

@levie Would be interesting to test Box files alongside CRM and billing records. Jylus compiles ingested records into a source-backed Context Pack. How much of that cross-source assembly do you leave to the agent today?

@hwchase17 Routing is one lever. Another is how much work reaches the model. With Jylus we prepare a bounded, source-backed Context Pack first. I'd test both together and compare cost per successful task, including retries.

@getcompoundai For long-running analysis, I'd want each result tied to the exact records it used. Jylus returns a bounded Context Pack plus source IDs for inspection. Can Compound consume that through MCP alongside its other data sources?

@hone The procurement use case caught my eye. Can a reviewer inspect the contract, invoice and supplier records behind each proposed saving? Jylus prepares that evidence before the model reasons. Would be useful to compare on a messy real case.

Try it with your own records. Ask a question the data cannot answer, then inspect the missing-evidence fields in the Context Pack: jylus.ai/try⁠�

Can your AI tell “false” from “we don’t have the evidence”? No approval record doesn’t mean rejected. No incident report doesn’t mean nothing happened. Jylus flags missing evidence in the Context Pack so the model can qualify its answer or ask for what’s missing.

@Yukiiiiiya Exactly. The approval can remain valid while the conditions behind it stop holding. That’s a stronger Jylus test: keep approval unchanged, update the linked supplier risk, then check what evidence reaches the agent before execution. Got a concrete example we can run?

9:00 — Approved. 9:07 — Approval revoked. 9:10 — Your AI agent executes. Which version of reality did it use? Yesterday, OpenAI shared its work with Ironclad on training AI agents to handle complex contracting workflows, including agreements, approvals and business rules. That raises a question every team deploying agents should ask: What happens when the evidence changes halfway through the job? Imagine an agent processing a supplier payment. The invoice is genuine. The original approval is genuine. The agent can cite both. But the supplier was placed on hold three minutes ago. Every document it retrieved could be authentic. The action could still be wrong. Before an agent acts, it needs to establish: • What is valid at the time of the decision? • Has anything superseded the original approval? • Are there conflicting records? • What evidence supports proceeding? This is what Jylus does. It resolves changing state and relationships, flags conflicts and missing evidence, and compiles source-backed context for the model to reason from. As agents gain the ability to move money, change access and update systems, the age of their evidence becomes a business issue. Here’s a useful test for your next agent demo: Change an approval halfway through the task. Does the agent catch it? Try Jylus with your own data: jylus.ai/try #AgenticAI #EnterpriseAI #Jylus

Why pay your AI to read 205,129 tokens when 2,532 scored higher? Our frozen 528-question test: Full context: 78.79% strict accuracy Jylus: 100% observed Same model. Same settings. Your hardest question. Your data. Try it: jylus.ai/try⁠�

The glue is usually the missing step between retrieve and generate. Separate retriever only helps if something then drops the wrong doc instead of passing it through. We do that as a pack: current record in, old one kept as history, proof id on what the model is allowed to read. If you have one question where it keeps opening the cousin document, I can show you what that pack looks like on your data. jylus.ai/try

@Nitijgjvk Citations do not settle a conflict. Hybrid search can rank last year’s policy first and the citation will still look clean. The loop is not the problem. Unresolved version is.

@Prayag12B The million-token window fits both versions. That is the failure, not the fix. Synthesis across 40 sections gets worse when two of them disagree and nothing has marked which one is current. Filter to the valid state, then reason.

@MemOS_dev Useful is the hard word. A fact that changed yesterday is still similar, so retrieval will keep serving it. Memory that cannot mark the old one as history will stay fluent and wrong.

@GauravGoyalAI That’s the distinction we built Jylus around. The tricky case is a correction received today that applies to last month. “What happened then?” and “What did we know then?” need different evidence. Do you track effective time and recorded time separately?

@nobeaten Stuffing the document is how both versions get in. The model then cites the one that reads better, not the one that is current. Cutting tokens only helps if the cut keeps the valid state.

@trytruvyx Illegitimate success is the right name. Faithful to the chunk, wrong on authority. The trace has to show which source was admitted, not only that the sentence matched the packet.