liveBook · Manning

Context isolation

As explained in Context Engineering

Separating work into distinct context windows so unrelated or competing information does not occupy the same working set.

Context isolation separates work into bounded context windows so unrelated or competing information does not occupy the same working set. Each unit receives only the instructions, tools, data, and state needed for its objective. Isolation can separate subtasks, authority levels, tools, or evidence streams. section 7.5 section 7.10

Unlike selection, writing, and compression, isolation restructures work across execution contexts. A coordinator decomposes the task, orders dependencies, merges results, and makes final decisions. Bounded units, typically subagents, do local work and return compact results; their detailed exploration, tool calls, failures, and evidence stay outside the parent context. section 7.5 section 7.1

Research-to-writer handoff

In a sequential CrewAI workflow, a researcher identifies trends and a writer uses the research task’s output to draft a post. The writer receives that output, not the researcher’s own reasoning. The task context argument defines the handoff boundary. section 7.9.4

Listing 7.5 Defining a collaborative crew with CrewAI
from crewai import Agent, Task, Crew, Process

researcher = Agent(role='Researcher', goal='Analyze trends',
backstory='Expert analyst') #1
writer = Agent(role='Writer', goal='Write post',
backstory='Content specialist')

research_task = Task(description='Identify 3 trends', agent=researcher)
writing_task = Task( #2
    description='Write post using research',
    agent=writer,
    context=[research_task] #3
)

crew = Crew(agents=[researcher, writer], tasks=[research_task,
writing_task], process=Process.sequential)
result = crew.kickoff()

The listing defines separate agents and uses the writing task’s context argument to pass the research result. section 7.9.4

The handoff must carry what the next unit needs: intermediate work stays local rather than automatically entering the receiving context. For long-horizon tasks, DeepAgents combines subagent isolation with a scratchpad and filesystem-based context writing. section 7.5 section 7.9.5

How Context Engineering covers it

How deep it goes
Explains coordinator responsibilities and subagent result boundaries, then demonstrates an explicit task handoff.
Examples
Uses a researcher-to-writer pipeline and a long-horizon agent workflow with a scratchpad and virtual filesystem.
What you can do after
Apply it via CrewAI by defining specialized agents and passing a prior task’s output through the receiving task’s context.
How it connects
Connects context isolation with context selection, context writing, context compression, and multi-agent orchestration.

Also explained in

Compare how 2 books explain context isolation →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Context Engineering. The section links point to the parts of the book it draws on.