Context Engineering Explained: What Your AI Agent Actually Sees @oracledevs
Context Engineering Explained: What Your AI Agent Actually Sees  @oracledevs
Uploaded November 2026 | Updated November 2026, 3 days ago
Read more here: blogs.oracle.com/developers/what-is-context-engineering-the-four-operations-that-decide-whether-your-agent-survives-production?source=:so:yt:or:awr:ot:::W7yt&SC=:so:yt:or:awr:ot:::W7yt&pcode=

What is context engineering? Context engineering is deciding what information reaches an AI model, in what form, as an agent works through a task: instructions, the user's request, relevant records, tool results and the conversation you keep.

Presented by Nacho Martínez, Data Scientist Advocate at Oracle. Nacho on LinkedIn: linkedin.com/in/jasperan

0:00 The missing rule
0:43 What context engineering means
1:22 What compaction can lose
2:13 Assemble the next input
3:10 Context and enforcement
3:49 Inspect your own agent

In this Oracle Developer Explains episode, Nacho Martínez starts with an illustration. A customer service agent is given a rule: refunds above two hundred pounds need a manager's approval. At first it follows the rule. After a long conversation, it proposes a nine-hundred-pound refund without asking anyone. The rule still exists; it just didn't reach this decision.

A model call doesn't automatically include everything your application has stored. Your application assembles the input for each call, and a tool result can become part of the next input. In the refund example, the application replaces older messages with a summary to make room, a process often called compaction, and the summary keeps the customer's problem but drops the approval rule. A bigger context window gives more room, but you still have to decide what belongs in it.

The fix rebuilds the next input: include the current approval policy from a maintained source, fetch the order facts when they're needed (and ask or stop if the lookup fails), and shorten the history while keeping the unresolved request. Including a rule doesn't guarantee a model will follow it, so the refund tool must enforce the approval too: context and enforcement are separate jobs. Retrieval is one part of context engineering, and a database can keep a fact after the conversation ends, but your application still has to select it.

When your agent gets something wrong, inspect that turn. Did the information exist? Was the correct version retrieved? Did it actually reach the model?

The refund example at 1:22 comes from the Oracle Developers article What is context engineering? The four operations that decide whether your agent survives production, which walks through the same failure and the choices behind it: blogs.oracle.com/developers/what-is-context-engineering-the-four-operations-that-decide-whether-your-agent-survives-production?source=:so:yt:or:awr:ot:::W7yt&SC=:so:yt:or:awr:ot:::W7yt&pcode=

The previous episode, What Is an Agent Harness?, is on the Oracle Developers channel. youtube.com/watch?v=9WtkwbkgQts&list=PLZsFfzJ3PRV4&index=3

Read the context engineering article, then use the three questions on one failed turn in your own agent. Subscribe to Oracle Developers for the next Explained episode.

#ContextEngineering #AIAgents #OracleAI
Context Engineering Explained: What Your AI Agent Actually SeesHow to Add Live Speech-to-Text to Low-Code ApplicationsKubernetes is complex. So are planes. | Kunal Kushwaha | Oracle Creator LabsTaking a Look at the new DBMS_OGA package in Oracle AI Database 26aiHow Frontier AI Models Are Changing Oracle Applications SecurityOCI Resource Analytics: Query and Visualize Your OCI Resource InventoryHow to Add Vector Search to Open WebUI with LiteLLMAndrew Ng Answers: What’s Overhyped in Agentic AI?Store Embeddings in the Database: One Transaction Moves BothOracle Developer Coaching: August News EditionBeyond the Context Window: Building Memory-Aware AI Agents with MCP and Oracle AI DatabasePart 2 - Setup CD3 toolkit in a single click using OCI Resource Manager
Oracle Developers |

Context Engineering Explained: What Your AI Agent Actually Sees

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER