AI_PLUMBER
SYSTEM_INDEX
UPLINK STATUS: OPTIMIZED
ACCESS_LEVEL: ADMIN_ROOT
SESSION_ID: 0x99_PIPE_FLOW
LAST_SYNC: 22.03.2026_04:00_GMT
©2026 AI_PLUMBER_CORP
architecture AI PLUMBER
Home / Wiki / Context Debt
Wiki · Core Concept

Context Debt

The accumulated cost of undocumented architectural decisions in AI systems.

Definition

Context debt is the accumulated cost of undocumented architectural decisions, tribal knowledge dependencies, and missing decision records in AI systems. Like technical debt, it compounds over time. Unlike technical debt, it is invisible until it causes a failure — a compliance question nobody can answer, a model behavior nobody can explain, or an infrastructure choice nobody remembers making.

How Context Debt Accumulates

Context debt grows through five primary mechanisms:

Implicit Decisions

Decisions made during development that are never recorded. Why was this model chosen over alternatives? Why was this threshold set to 0.85? Why does this agent have write access? If the answer lives only in someone’s head, it is context debt.

Tribal Knowledge

Operational knowledge held by individuals rather than systems. When the person who configured the model routing leaves, the routing logic becomes an undocumented black box. The system still works — until it doesn’t, and nobody knows why.

Missing Architecture Decision Records

ADRs are the minimum viable governance for architectural decisions. Without them, every future discussion about “why we did it this way” starts from zero. Teams re-litigate decisions that were already made, wasting cycles and risking contradictory choices.

Undocumented Constraints

Regulatory constraints, business rules, and safety boundaries that are enforced in code but never documented as governance policy. When the code changes, the constraint disappears — and nobody notices until the compliance audit.

Configuration Drift

Production configurations that diverge from documentation over time. Kill thresholds adjusted in production but not in the governance spec. Agent permissions expanded for debugging and never reverted. Each drift is a context debt increment.

Symptoms

You have context debt when:

Why AI Systems Are Especially Vulnerable

AI systems accumulate context debt faster than traditional software because they involve more implicit decisions. Model selection, training data choices, feature engineering decisions, threshold calibration, agent permission scoping, cost guardrail settings — each is an architectural decision that shapes system behavior. Traditional software has well-established documentation patterns. AI systems often don’t. The result: AI systems carry more context debt per line of code than any other type of software.

Mitigation Strategies

01Architecture Decision Records (ADRs) — document every significant decision with context, alternatives considered, and rationale. ADRs are the minimum viable governance for context debt prevention.
02Decision logs in governance infrastructure — make decision recording a system feature, not a process burden. The AI Plumber framework’s attributable actions pattern automates this.
03Configuration-as-code with version control — all governance configurations (thresholds, permissions, routing rules) stored in version control with commit messages explaining changes.
04Regular context audits — quarterly reviews of documentation freshness. If any documentation is older than the code it describes, it is context debt.
05Onboarding as a canary — how long it takes a new team member to answer “why does this system work this way?” is a direct measure of context debt.

The Cost of Ignoring Context Debt

Context debt has a compounding interest rate. Early in a project, the cost is low — the team remembers decisions, the system is small enough to hold in one person’s head. As the system grows, as team members change, as regulatory requirements evolve, the cost accelerates. By the time context debt causes a compliance failure or a production incident, the remediation cost is orders of magnitude higher than the documentation cost would have been. The cheapest time to document a decision is when you make it.

Concerned about context debt in your AI systems?

Book Architecture Review →