Problem
Most enterprise AI errors don’t happen because the model is incapable. They happen because the context arrives wrong: stale, incomplete, duplicated, without permissions, without an owner, or mixed with unapproved information.
Companies usually treat context as “available documents.” For agents, that is insufficient. Context is operational input.
If you feed the machine poorly, you don’t get intelligence. You get false confidence.
Thesis
Context Supply Chain is the chain that governs where corporate knowledge comes from, how it is transformed, and when it can be used.
RAG alone isn’t enough. You need a supply chain: authorized sources, permissions, versioning, freshness, chunking, ranking, memory, traces, and verification against outcome.
Framework
The chain has eight stages:
- Source: where the information originates.
- Ownership: who is responsible for its quality.
- Permission: who can see it.
- Freshness: how long before it expires.
- Packaging: how it is fragmented and labeled.
- Retrieval: how it is selected.
- Use: how it enters a decision.
- Feedback: which errors flow back to the source.
Mini-case: a sales agent answers pricing questions. If it retrieves an old proposal, a public page, and a contradictory internal note, the output may sound confident and be wrong. The solution isn’t “better prompt.” It’s a governed context chain.
Measurable signal: percentage of sources used by agents that have an owner, permission, and expiration date.
Position: context without governance is data debt with a friendly interface.
Why it matters now
MCP formalizes how to expose tools and resources to models. LangGraph and LangSmith push memory, state, traces, and evaluation as pieces of agentic systems. OpenAI recommends building agents around instructions, tools, guardrails, and evaluation, not just around the model.
The focus shifts from “which model do we choose” to “what the system knows, with what permission, and with what evidence.”
Anti-example
“We index the entire Drive.”
That is not a context strategy. It’s massive ingestion. It can mix drafts, old PDFs, sensitive data, duplicates, and documents without authority.
Protocol (3 steps)
- Classify sources by authority. Official, work in progress, historical, private, external.
- Define expiration. Pricing, policies, legal, and product information cannot live indefinitely.
- Close the loop. Every incorrect answer must point to the source, retrieval, or freshness.
| Stage | Control | Question |
|---|---|---|
| source | authority | who approves it |
| permission | ACL | who can see it |
| freshness | expiration | is it still valid |
| retrieval | ranking | why was it selected |
| feedback | correction | what is learned |
Related
- Context Architecture: designing memory and knowledge for AI systems
- Enterprise AI Search: why internal search is becoming an operating system
- Context Budgeting: saving tokens without blinding the agent
Sources consulted
Next step
Choose a critical response that an agent would give today. Trace the necessary sources and mark which is official, who maintains it, when it expires, and how you know it was used.
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.