.svg){kind=logo}
Retrieval
Pull only what the current step needs, using search, filters and entity lookups rather than dumping whole corpora.
Context engineering · The definitive guide
Context engineering: what AI agents need to know, at every step.
Better prompts stopped being enough the moment AI started acting. This guide covers what context engineering is, how it differs from prompt engineering and RAG, where it breaks, and why in the enterprise it is really a data and governance problem.
- Seven components of context
- Four failure modes
- Six techniques that work
Context window · step 4 of 7
Assembled
TaskApprove vendor invoice #INV-2291
Input
Instructionsrole, goal, output format
system
Business entitiesvendor, PO, contract
graph
Decision history3 similar approvals
traces
Policyauto-approve under limit, 3-way match
control
Toolserp.match_invoice, erp.approve
mcp
Pruned
Excluded: 41 unrelated emails, stale price list.
Illustrative example. The right context is as much about what is left out as what goes in.
Works with
.svg)
Why context matters
Better prompts stopped being enough. Context is now the risk.
Bigger context windows did not remove the problem. They made it easier to fill the window with things that confuse the model.
A model only knows what is in front of it. As agents run longer, call tools and act on real systems, what they see at each step matters more than how any single prompt is worded.
-
01
Context rot
As the window fills over a long session, the model attends less reliably to what matters.
-
02
Overload
Too many documents or tools, so the relevant signal is buried and tool choice gets worse.
-
03
Poisoning
A hallucination, bad tool result or injected instruction enters context and is treated as fact.
-
04
Conflicting and stale data
Two systems disagree, or the context was right last month and the agent acts on yesterday’s contract or last quarter’s policy.
What is context engineering?
Six techniques that keep context sharp
Context engineering is the discipline of designing what information, memory, tools and constraints an AI agent receives at each step of a task, so it can reason and act correctly within a limited budget of attention.
It selects, shapes and orders seven things: instructions, retrieved knowledge, memory, tool definitions, state, policies and examples.
Compaction and summarisation
Condense long histories and tool outputs into the facts that still matter, so the window stays focused.
Structured memory
Store notes, decisions and state outside the window in a structured form the agent can read back when needed.
Sub-agents
Give focused sub-tasks to agents with their own clean context, and return only their conclusions.
Just-in-time loading
Keep references such as IDs and paths in context, and load the full content only when the agent decides it needs it.
Ranking and ordering
Rank candidates by relevance, freshness and authority, and place the most important context where the model weighs it most.
How Context OS works
From scattered systems to governed context
Context OS is the governed context layer between your systems and your agents. Decision traces close the loop, so every decision becomes context for the next one.
Connect
Unify data from ERP, CRM, data platforms and documents with connectors and entity resolution.
Model
Build business context: a knowledge graph, shared definitions and decision history.
Govern
Policy and permissions decide who may see and do what, before context reaches the agent.
Assemble
Retrieve, rank and compact context for each step, then record every decision as a trace that feeds back as history.
What’s different
Prompts and RAG improve answers. Context engineering makes agents decision-ready.
Prompt engineering and RAG are parts of context engineering, not alternatives to it.
| Prompt engineering and RAG alone | Context engineering with Context OS | |
|---|---|---|
| Focus | How one instruction is worded, or which documents are fetched for a query | Everything the model sees, across every step |
| Scope | One call or one query | Instructions, knowledge, memory, tools, state and policy |
| Time horizon | Static, or per query | Dynamic, across long-running agent sessions |
| Main failure | Ambiguous instructions, irrelevant or missing passages | Wrong, stale, conflicting or excessive context, which we design against |
| Who owns it | Prompt author or search engineer | AI engineers with data, security and business owners |
| Best for | Chat, simple generation and question answering over documents | Agents that plan, use tools and act |
What agents draw on
Four sources of governed context
In the enterprise, context engineering is a data problem. Techniques inside the agent cannot fix context that is wrong at the source.
01
Business entities
Customers, vendors, assets, contracts and accounts, resolved across systems into one view.
Unify Data 02Knowledge graph
The relationships between entities, so agents can follow how things connect.
Enterprise knowledge graph 03Decision history
How similar decisions were made, by whom and with what outcome, as precedent.
Decision traces 04Policies and permissions
The rules and access rights applied before context reaches the agent and before it acts.
Governed actionsResults
Built for production from week one
Figures as published on elixirdata.co.
FAQ
Context engineering, answered
Straight answers on what context engineering is, how it differs from prompting and RAG, and why enterprises need a governed context layer.
Designing what information, memory, tools and constraints an AI model or agent receives at each step, so it can reason and act correctly within a limited context window.
Prompt engineering is about wording a single instruction. Context engineering covers everything the model sees across a task, including retrieved knowledge, memory, tool definitions, state and policies.
No. Retrieval-augmented generation is one technique within context engineering. Agents also need memory, tools, state and constraints, which retrieval alone does not provide.
Agents run many steps, call tools and act on real systems. Wrong, stale or excessive context at any step leads to wrong actions, so what the agent sees matters more than how it is prompted.
No. Larger windows hold more, but models still attend less reliably as the window fills, and more content means more chances for conflicts and noise. Selection still matters.
The Model Context Protocol standardises how agents reach tools and data sources. It is the plumbing; context engineering decides what flows through it and when.
A governed layer between enterprise systems and AI agents that unifies data, models business entities and relationships, keeps facts current, applies permissions and serves the right context to each agent.
Context OS gives every agent governed, current business context: unified data, a business knowledge graph, decision history and policy checks, with decision traces that feed each outcome back as context.
Give your agents the context to decide well
In one session, we map the context behind your first three agent decisions: where it lives, where it conflicts, and what Context OS would serve.