Demand-Driven Context Management for Enterprise AI Agents

ikea · E-commerce · 2026 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise agent context management

IKEA's delivery and services domain needed AI agents to complete enterprise work items that depend on undocumented institutional knowledge, not just general coding knowledge or existing documentation indexed through RAG/MCP.

Constraints:

• The domain has more than 100 engineers across six product teams.
• Enterprise institutional knowledge is partly outdated, unreliable, duplicated, or tribal.
• RAG and MCP layers fail when the underlying knowledge base is incomplete.
• Agents need task-relevant context without being forced to consume the whole knowledge monolith.
• The approach has to improve through incident/task cycles instead of one-time documentation pushes.

Architecture

IKEA's demand-driven context approach inserts a context-block layer between monolithic enterprise knowledge and retrieval/MCP layers. Agents are assigned real incidents or work items, fail or ask questions where institutional knowledge is missing, pull answers from humans and existing systems, and document those answers as reusable context blocks so future agents can resolve similar tasks with less discovery.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  work-item-queue["Work item queue"] --> agent-worker["Agent worker"]
  agent-worker --> retrieval-and-mcp-layer["Retrieval and MCP layer"]
  retrieval-and-mcp-layer --> knowledge-monolith["Knowledge monolith"]
  knowledge-monolith --> human-knowledge-holders["Human knowledge holders"]
  human-knowledge-holders --> agent-worker
  agent-worker --> context-block-store["Context block store"]
  context-block-store --> agent-worker
  agent-worker --> confidence-tracker["Confidence tracker"]
  context-block-store --> confidence-tracker

Components

Data flow

A work item or incident is given to an agent. The agent uses existing retrieval and MCP servers, identifies missing institutional context, asks questions or receives human answers, converts the discovered knowledge into a context block, and reuses the accumulated blocks in later incidents. Repeated cycles raise confidence as tribal knowledge is converted into curated, reusable context.

Why this architecture

The talk argues that pushing all documentation into retrieval fails because the knowledge base itself is not reliable. A pull-based context-block layer mirrors human onboarding: assign real work, discover gaps, and document the useful 20% as reusable context.

Models used: Coding agents · LLM agents · RAG/MCP-backed agents

Scale

• The presenter works in a delivery and services domain with more than 100 engineers and six product teams.
  “we are like almost more than 100 engineers and six Pedics teams all together. It's like a mini company within the company itself.”transcript, stated
• The talk cites McKinsey's 88% AI-use figure and 6% value-creation figure as the macro problem.
  “it's from the McKinsey this year, like 88% of all companies use AI, but they only see like uh 6% of value creation.”transcript, stated
• The presenter estimates 40% of institutional knowledge is tribal knowledge.
  “the major problem is 40% of the knowledge is always uh tribal knowledge, which means people know how things work.”transcript, stated
• Across 14 incidents, confidence rose from 1.5 to 4.4 in the example.
  “on the 14 incidents, it basically actually able to go to a confidence level of 4.4”transcript, stated

Alternatives considered (roads not taken)

Build more MCP servers over existing documents

Rejected because: The presenter says MCP/RAG accuracy is limited when the underlying institutional knowledge is missing or bad.

Push all company knowledge to agents upfront

Rejected because: The approach instead assigns work and lets agents pull the knowledge they actually need.

Give agents 100% of documentation

Rejected because: The presenter argues the useful 20% should become context blocks and the rest can stay linked for lookup.

Patterns

agent-memory (context)

Agents convert discovered institutional knowledge into reusable context blocks for later tasks.

rag-retrieval (retrieval)

The approach positions context blocks between the enterprise knowledge monolith and existing RAG/MCP retrieval layers.

data-quality-loop (knowledge)

Incident cycles reveal missing and unreliable knowledge, then improve future context quality.

confidence-thresholding (evaluation)

The example tracks confidence levels across incidents as context blocks accumulate.

Tradeoffs

Use demand-driven context

Chose pull-based context discovery from real work over push-based all-document onboarding.

Rationale: The agent learns the institutional knowledge it actually needs for tasks, similar to human onboarding.

Cache the useful context

Chose small curated context blocks plus links over feeding the whole documentation corpus.

Rationale: The presenter frames useful knowledge as an 80/20 problem.

Gotchas (what bit them)

Retrieval layers cannot fix missing institutional knowledge.

Root cause: Large portions of knowledge are outdated, unreliable, duplicated, or purely tribal.

Resolution: Use task-driven cycles to discover and document the missing context as blocks.

More MCP servers can increase surface area without improving agent outcomes.

Root cause: Servers point at the same flawed knowledge monolith.

Resolution: Break institutional knowledge into task-proven context blocks instead of only adding connectors.

Evaluation

The talk evaluates the approach through incident cycles, confidence scores, missing-knowledge discovery, context-block reuse, and whether agents can resolve new incidents with less human intervention.

Metrics: Confidence score · Incident cycles · Missing context items · Context block reuse · Human answers required · Task resolution quality

Outcomes

• The approach surfaced undocumented knowledge through incident cycles.
• The example confidence level improved from 1.5 to 4.4 over 14 incidents.
• Agents documented answers and built reusable context blocks.
• The speaker published a preprint on demand-driven context in March 2026.

The case is grounded in an IKEA domain with more than 100 engineers and six product teams.