How We Reduced AI Agent Friction by Learning from Our Newest Agent

We audited 18 AI agents for operational friction and discovered something counterintuitive: our newest, simplest agent had the least friction of all. Not the most battle-tested one. Not the one with the most comprehensive documentation. The one we'd set up most recently, with the fewest rules.

This paradox—that mature agents accumulate friction while new agents start clean—changed how we think about AI agent governance entirely.

What We Mean by "Friction"

Agent friction is anything that slows an agent down, causes errors, or leads to wasted work. It's the accumulated weight of operational complexity—rules that contradict each other, documentation that's grown too long to process efficiently, database routing that requires tribal knowledge, and permission boundaries that aren't clearly defined.

We measured friction by analyzing the operational knowledge base entries associated with each agent. These entries represent hard-won lessons—things that went wrong and the solutions that were documented afterward. More entries doesn't necessarily mean more friction, but the pattern of entries tells a story.

The Five Friction Patterns

Across all 18 agents, five distinct friction patterns emerged:

The Reference Agent: What "Low Friction" Looks Like

Our newest agent stood out as the clear outlier. Set up recently for a focused project, it had:

This agent completed tasks faster, made fewer errors, and required less context per session than agents that had been running for months. The data was clear: simplicity was winning.

The Paradox of Maturity

Here's the uncomfortable truth: every incident response makes your agents a little slower. Every "never do X" rule adds weight. Every cross-reference adds complexity. The agents that have survived the most problems carry the most baggage.

It's the same pattern organizations see with legacy codebases. The system that's been in production for five years has more features, more edge-case handling, more defensive code—and is dramatically harder to work with than a fresh implementation that does 80% of what's needed with 20% of the complexity.

"Our most experienced agent had accumulated so many operational warnings that it spent more context tokens on 'don't do this' rules than on understanding the actual task."

The Remediation Plan

Using the reference agent as a template, we developed a systematic approach to reducing friction across the fleet:

1. Audit configuration length. Any configuration file over 150 lines gets a mandatory review. What can move to the shared knowledge base? What's defensive documentation that should be a gotcha entry instead of a hard rule?
2. Eliminate circular inheritance. Each agent should have exactly one source for each rule. If a rule appears in the team instructions, it doesn't need to also appear in the project configuration.
3. Clarify database routing. Every agent's configuration explicitly states which databases it connects to and why. No ambiguity, no "check the docs" references.
4. Streamline gatekeeper handoffs. Create clear specification templates so that when agents need changes from specialist agents, the format and expectations are standardized.
5. Separate hard rules from soft knowledge. Hard rules go in configuration files (always enforced). Soft knowledge goes in the searchable knowledge base (surfaced when relevant). Stop putting everything in one place.
6. Create a new-agent template. Based on the reference agent's structure, every new agent starts with a clean, minimal configuration that follows proven patterns.

Results So Far

The friction reduction effort is ongoing, but early results are encouraging:

• Three agent configurations reduced by 30-40% in length without losing any critical rules
• Circular documentation inheritance eliminated across the fleet
• A new agent template ensures future agents start with the reference pattern, not the legacy one
• Universal rules consolidated into a single team instructions file (see our companion post on building a team instructions layer)

Lessons for Your AI Operations

If you're running AI agents in production, the friction audit is one of the highest-leverage activities you can do. Here's how to start:

Measure configuration weight. How many tokens does each agent consume just reading its rules before it starts working? If the answer is "I don't know," that's your first finding.

Categorize your rules. What percentage are positive instructions ("do this") versus defensive warnings ("don't do that")? A high ratio of defensive rules suggests accumulated incident baggage.

Find your reference agent. Which agent has the cleanest configuration and the best performance? Study what it does differently. The answer is usually "less."

Schedule regular audits. Friction accumulates gradually. A quarterly review of agent configurations prevents the slow creep toward bloated, contradictory rule sets.

"The best documentation isn't the most comprehensive—it's the most focused. An agent with five clear rules outperforms one with fifty accumulated warnings."