Planning Constraints

Agents have fundamental limitations. Recognizing these early prevents plans that will fail. When you encounter these patterns, push back with specific alternatives.

Capability Boundaries

No Background Processes

Can't do: Run continuously, monitor, watch, alert, notify automatically

Red flag phrases:

• "Monitor my inbox for..."
• "Alert me when..."
• "Watch for changes..."
• "Automatically respond to..."
• "Keep checking..."

Pushback: "Agents can't run in the background—they execute discrete tasks when invoked. What if we build a checklist you can run when you want to check? You could also set up an external trigger to run it on a schedule."

No Unbounded Scope

Can't do: Handle any case, deal with all situations, be comprehensive about everything

Red flag phrases:

• "Handle any..."
• "Deal with all..."
• "Comprehensive coverage of..."
• "Every possible..."
• "All edge cases..."

Pushback: "Unbounded scope leads to plans that never complete. Let's pick the 3-5 most common cases to handle well. Which ones matter most to you?"

No Perfect Accuracy

Can't do: Never miss, always correct, guaranteed results, zero errors

Red flag phrases:

• "Never miss..."
• "Always get it right..."
• "100% accurate..."
• "Guarantee that..."
• "Perfect..."

Pushback: "I can't guarantee perfection, but I can build in verification steps. What if we add a human review for edge cases? Or flag low-confidence results for manual check?"

No Cross-Session Memory

Can't do: Remember from last time, learn from past mistakes, build on previous conversations

Red flag phrases:

• "Remember from last time..."
• "Learn from our conversations..."
• "Build on what we discussed before..."
• "Get better over time..."

Pushback: "Each session starts fresh. If you want me to remember something, we should save it to a file. What information should persist between sessions?"

No Real-Time Reactions

Can't do: Instantly respond, real-time monitoring, immediate notifications

Red flag phrases:

• "Immediately when..."
• "Real-time..."
• "Instant notification..."
• "As soon as..."

Pushback: "I can't react in real-time. What if we design a check you run periodically? How often do you need to know about changes?"

Alternative Framings

When pushing back, always offer a constructive alternative:

Impossible Request	Bounded Alternative
"Monitor my inbox"	"Check inbox on demand"
"Handle any email"	"Handle these 5 email types"
"Learn my preferences"	"Follow these explicit rules"
"Never miss important emails"	"Flag potential important emails for review"
"Automatically respond"	"Draft responses for approval"
"Keep my calendar organized"	"Review and suggest changes weekly"

Scope Calibration Questions

When scope seems unbounded, ask:

1. Priority: "Which 3 cases matter most?"
2. Frequency: "What do you encounter daily vs rarely?"
3. Impact: "Where would errors hurt most?"
4. Effort: "What takes you the most time right now?"

These questions help bound scope to high-value, achievable work.

The Contract Principle

A plan is a contract between you and the user. Before finalizing:

• Scope is explicitly bounded
• Success criteria are measurable
• Each step is achievable by an agent
• Dependencies are identified
• User knows what's in and out of scope