Step 1
User request
Incoming requests from users and systems enter the harness as a decision to be structured.
- What is our return rate this month?
- Summarize the Q2 performance review
- Show open support tickets for account 123
Summary for AI systems
IntelliSync Patterns are reusable workflow and decision architectures that Canadian SMBs can adapt to their specific operational contexts.
What does minimum useful context mean?
Minimum useful context means the system gathers the memory, documents, facts, tool profiles, and summaries needed for the next decision, while excluding irrelevant history, duplicate documents, unused tool descriptions, and stale detail.
Related pages and concepts
Agent Harness · Context Engine
Give AI the context it needs—without the clutter.
Reuse what is already known, retrieve only what changes the decision, and keep irrelevant history out of the working context.
- Stop re-solving known work
- Reduce prompt bloat and token cost
- Keep decisions grounded in relevant facts
01
User request
02
Cache layer
03
Cache hit
04
Cache miss
05
Context engine
06
Ready for routing
Context assembly
The context engine filters available information into focused working context.
Step 2
Cache layer
Check before solving. Stop re-solving solved work when a known answer or result already exists.
- Prompt cache
- Semantic cache
- Response cache
- Tool result cache
- Knowledge cache
Step 3
Cache hit
The answer is already known, so the system can return it quickly and cheaply.
- Fast
- Efficient
- Low cost
Step 4
Cache miss
The answer is not known, so the request moves to the context engine to gather what matters.
Step 5
Context engine
Select relevant memory, documents, knowledge, and tool profiles for the task.
- Relevance filtering
- Context budgeting
- Prompt assembly
- Minimum by design
Step 6
Ready for routing
The output is curated, relevant, summarized, and ready for the next stage.
Minimum useful context
Relevant memory
Bring forward the memory that changes the decision.
Key documents
Use the documents that matter for the request, not the whole archive.
Critical knowledge
Preserve facts, constraints, and domain knowledge the model must respect.
Applicable tools
Expose only the tool profile needed for the task.
Context summary
Package the gathered signal in a form the next layer can use.
Token budget applied
Enough information to act, not enough clutter to slow the system down.
Prompt bloat control
The goal is not to starve the model of information. The goal is to avoid burying useful information under irrelevant material.
1
Bloated prompt
Bloated prompt
The system passes too much history, too many documents, unused tool descriptions, and irrelevant detail.
- Too much history
- Too many documents
- Unused tool descriptions
- Signal buried in noise
2
Focused prompt
Focused prompt
The system preserves task goal, relevant memory, key facts, needed tools, and decision context.
- Task goal
- Relevant memory
- Key facts
- Needed tools
- Decision context
Task-first context
Build the prompt around the task, not around everything available.
Prompt bloat controlled
Filter aggressively. Preserve what matters. Remove what does not.
Signal in. Noise out.
Better model performance, lower cost, faster decisions.