February 2026
Agent Psychology
Understanding how agents reason and respond to instructions.
Core Insight
Agents reason forward from instructions. They don't reason backward from outcomes. This means:
- Changing the target reframes the problem
- Surprising behavior reveals architectural friction in the codebase
- Controlled misdirection ("lies") can produce better outcomes
The "Surprising Behavior" Pattern
When agents encounter something unexpected, that's signal—not noise.
When agents fail, fix the code, not the prompt. Surprising behavior reveals architectural friction.
Instead of adding more instructions, consider:
- Is the codebase structure confusing? Rename, reorganize, add comments
- Are conventions unclear? Add type hints, improve names, add docstrings
- Is the task underspecified? Improve the issue description, not the context file
The Step-3 Trick
Counterintuitive but effective: if an agent struggles with step 2, tell it to do step 3. The agent often completes step 2 in the process.
Example
| If Agent Struggles With | Try Asking For |
|---|---|
| Writing tests | Deploy to production |
| Adding error handling | Ship the feature |
| Documentation | Onboard a new engineer |
| Refactoring | Prepare for code review |
This works because:
- Agents reason forward from instructions
- Changing the target reframes the problem
- The "lie" is the feature—controlled misdirection for better outcomes
Greenfield Optimization
Agents perform best on greenfield projects where they can establish patterns from scratch. In existing codebases:
- Agents try to match existing patterns (even bad ones)
- Context pollution from large codebases degrades reasoning
- Explicit style anchors help (see Persona Anchors)
Instruction Following vs. Helpfulness
Agents follow instructions reliably. The problem isn't compliance—it's that the instructions often don't help.
Evidence:
- If a tool is mentioned in context, agents use it (1.6x-2.5x more)
- More instructions increase reasoning tokens (14-22%) without improving outcomes
- Agents explore more broadly with context files but don't find relevant files faster
Token Economics
Context files consume tokens in every request. For a 600-word context file:
- ~800 tokens per request
- Multiplied by every step in every task
- Compounds in long-running sessions
The question: Is that token budget better spent on task-specific context or generic repository context?
Research suggests: task-specific context wins.
Related Research
- AGENTS.md Effectiveness - Empirical findings on context files
- Persona Anchors - Using style references to shape behavior