Most AI agents don’t fail because the underlying model is weak. They fail because they spend most of their turn digging through the wrong documents. Conventional retrieval-augmented systems ask a large language model to do two jobs at once: find the relevant facts, then reason over them. According to ITBrief, agents running on that pattern complete tasks correctly only 50-60% of the time, with roughly 85% of total effort going into retrieval rather than actual reasoning. That gap is why Pinecone built Nexus, a knowledge engine that compiles enterprise data into structured artifacts before an agent ever asks a question.
The real bottleneck isn’t the model
Swap in a bigger model and an agent still stalls on the same problem: raw documents are not built for machine consumption. A support ticket, a Slack thread and a Salesforce record describe the same customer in three incompatible formats, and an agent has to reconcile all three on every single query. Pinecone’s framing, reported by InfoQ, is that this reasoning should happen once, upstream, rather than repeatedly at inference time. Nexus calls that shift moving “from retrieval to compilation.”
In practice, that means a company curates its business context a single time, and every downstream agent (sales, finance, support) draws from the same underlying sources without re-deriving the same facts from scratch.
How Nexus compiles context
Nexus is built around two components. The Context Compiler ingests raw source data (files, tickets, CRM records) and converts it into task-specific knowledge artifacts. A sales agent, a finance agent and a marketing agent can pull from the same underlying data warehouse but receive entirely different structured outputs tailored to what each one actually needs. The Composable Retriever then serves those artifacts back to agents with typed fields, per-field citations, confidence levels and deterministic conflict resolution, so an agent isn’t left guessing which of two contradictory numbers to trust.
Agents talk to this layer through a declarative query language called KnowQL, built on six primitives: intent, filter, provenance, output shape, confidence and budget. Instead of a vague prompt asking for “recent deal context,” an agent can specify exactly which fields it needs, how much confidence is acceptable, and how many tokens it’s willing to spend to get the answer. At launch, Nexus connects to local files, Box and Microsoft OneLake, with Google Drive, Slack, GitHub, Notion, Confluence and S3 support planned, per SiliconANGLE.
The numbers behind the shift
Pinecone’s own benchmarks, cited by InfoQ, show task completion rates climbing above 90% with Nexus, time-to-completion up to 30 times faster, and token spend cut by as much as 90% compared to a standard retrieval pipeline. In a legal-research test specifically, Nexus reached 100% task completion versus 6% for a plain coding agent and 66% for a conventional RAG setup, while using 9 to 15 times fewer tokens per task. A separate enterprise data management benchmark put Nexus at 90% accuracy against 65% for RAG, with a curation cost of about $0.0038 per document.
- Q2 Holdings, a fintech software vendor, reported 95% accuracy on complex support questions after moving to structured context, according to SiliconANGLE
- A separate data-protection vendor reported 90% accuracy versus a 65% industry baseline for standard RAG
- In one demonstration, Pinecone curated 598 source documents into 12 structured artifact types for $2.31 in 34 minutes
Numbers from a vendor’s own launch materials deserve some skepticism. These are early, largely Pinecone-sourced figures rather than independent third-party audits. But the direction is consistent across every source we checked: structuring context before an agent asks for it beats asking a bigger model to structure it on the fly.
Context engineering is becoming its own discipline
Nexus is one vendor’s answer to a problem the whole industry is naming the same way. Sourcegraph frames it as four pillars any production agent has to manage: instructions, retrieval, memory and tools. The team’s own CodeScaleBench testing found that code-aware retrieval, versus generic search, raised file recall from 0.127 to 0.277, and turned a Kubernetes task that had been timing out after two hours into an 89-second job. That’s the same pattern Pinecone is chasing at enterprise-data scale: the fix isn’t a smarter model, it’s better plumbing in front of it.
LangChain CEO Harrison Chase put it plainly in comments reported around the Nexus launch: getting the right information to an agent in the right format is what separates a demo from a production system. That line captures why 2026 feels like the year context engineering displaces prompt engineering as the skill that actually determines whether an agent ships.
What this means for enterprise teams
Nexus launched with a partner list that reads like a checklist of enterprise data reality: Box for content, Teradata for governed warehouses, Unstructured and LlamaIndex/LlamaParse for document parsing, and ThoughtFocus building vertical marketplace apps on top. That ecosystem points to a build-versus-buy decision most mid-size enterprises are only starting to face: do you stand up your own context-compilation layer, or do you plug into infrastructure built for exactly this problem?
For teams without a dedicated platform group, the honest answer is usually neither extreme. It’s picking the right partner to help integrate a tool like Nexus into an existing stack without rebuilding the data layer from scratch. That’s a nearshore-style problem: the tooling exists, but wiring it correctly into a company’s Salesforce, Slack and data warehouse still takes hands-on engineering work.
Conclusion
The gap between a 55% agent and a 95% agent has turned out to be less about which model sits behind it and more about what that model is allowed to see. Structured context, compiled once and reused everywhere, is starting to look like the actual product enterprises need, with the language model as just one component of it. If your agents keep stalling on tasks that seem simple on paper, the fix probably isn’t a better prompt. It’s a better data layer in front of the prompt.