The Indic Pacific Glossary

Agent debt is the compounded cost of rapidly prototyping autonomous AI workflows without establishing rigorous architectural hygiene. It is a modern, AI-specific evolution of technical debt.

When developers build agentic systems prioritizing speed over structure, they accumulate hidden complexities. Unlike traditional software where technical debt results in buggy but deterministic code, agent debt results in emergent, unpredictable, and degrading behaviour over time.

Here is a breakdown of how agent debt accumulates and why it is so difficult to manage.

Core Drivers of Agent Debt

• Prompt Collision: Adding "band-aid" instructions (e.g., "Never do X," "Always do Y") to fix immediate edge cases. Over time, these stacked system prompts conflict, causing the agent to become paralyzed or behave erratically because it cannot satisfy all constraints simultaneously.

• Memory Pollution: Failing to curate what the agent remembers. When raw logs, irrelevant conversational tangents, and outdated facts are dumped into the context window, the agent's reasoning degrades. It starts prioritizing noisy past data over current instructions.

• Tool Overlap: Providing an agent with multiple functions or APIs without clear, mutually exclusive descriptions. The agent wastes compute trying to figure out which tool to use, gets stuck in infinite loops, or executes the wrong action entirely.

• Context Entanglement: Building workflows where the output of one poorly defined agentic task feeds directly into another. If the first agent slightly alters its formatting or reasoning, the downstream agent silently fails or hallucinates.

Why Agent Debt is More Dangerous Than Technical Debt

Traditional technical debt is relatively straightforward to debug: a poorly written function will eventually throw an error, and a stack trace will point you to the exact line of code that broke.

Agent debt is insidious because the system rarely crashes outright. Instead, it "fails silently" by producing slightly worse logic, taking longer to execute, or doing "weird things." Because the logic is probabilistic rather than deterministic, diagnosing the root cause—whether it was a polluted memory vector, a conflicting system prompt, or a misunderstood tool description—becomes a massive forensic undertaking.

The "High-Interest Credit Card"

Gary Marcus's reference to the 2014 Google paper, "Machine Learning: The High-Interest Credit Card of Technical Debt," perfectly contextualizes this. The paper argued that it is remarkably easy to build ML models fast, but massively expensive to maintain them because everything is entangled—changing one input changes the entire system's behaviour.

Agent debt is the ultimate high-interest credit card. You get an impressive demo working in a weekend, but six months later, the system is a fragile house of cards, and the interest payments—the time spent debugging hallucinations and wrestling with prompt conflicts—bankrupt the engineering team's velocity.