Skip to main content

Failure Modes & Drift Taxonomy

Agentic development introduces failure modes that often:

  • pass basic tests
  • “work” in the short term
  • surface only after drift accumulates

This taxonomy is a practical map of the failures ADG is designed to prevent.

Category 1 — Boundary Failures​

1.1 Tenancy leakage​

Symptoms:

  • cross-tenant reads or writes
  • inconsistent tenant scoping in queries
  • “admin” paths that bypass tenancy inadvertently

Typical causes:

  • missing tenant filters
  • RLS policies absent/incomplete
  • bypass roles (owner roles, misconfigured DB roles)
  • background tasks executed without correct tenant context

1.2 Authorization ambiguity​

Symptoms:

  • endpoints that succeed without proper permission checks
  • “helper” routes not protected
  • inconsistent admin checks across routes

Typical causes:

  • duplicated auth logic
  • inconsistent middleware usage
  • agent-generated endpoints that skip guardrails

Category 2 — Coherence Failures​

2.1 Architectural intent drift​

Symptoms:

  • new patterns appear without explanation
  • ad hoc abstractions multiply
  • duplicate “ways to do the same thing”

Typical causes:

  • agents optimize locally
  • humans approve diffs without full context
  • no enforced architectural constraints

2.2 Documentation desynchronization​

Symptoms:

  • docs describe behavior that no longer exists
  • onboarding docs become untrustworthy
  • “tribal knowledge” replaces documentation

Typical causes:

  • docs treated as optional
  • agents update code faster than docs
  • no doc sync gates

Category 3 — Integration & Environment Failures​

3.1 Hidden coupling​

Symptoms:

  • changes in one subsystem break another unexpectedly
  • env vars renamed, shared, or collapsed
  • staging/prod boundaries blur

Typical causes:

  • agent refactors config without semantic understanding
  • no environment isolation tests
  • insufficient “blast radius” analysis

3.2 Schema drift and migration damage​

Symptoms:

  • migrations that work in dev but fail in prod
  • orphaned migrations
  • inconsistent naming and constraints
  • runtime code expecting schema that doesn’t exist

Typical causes:

  • agents generate migrations without lifecycle awareness
  • missing schema diff checks
  • insufficient migration verification in CI

Category 4 — Observability Failures​

4.1 Non-replayable actions​

Symptoms:

  • you cannot reconstruct what the agent did
  • logs lack correlation IDs or context
  • “it changed something” without full trace

Typical causes:

  • logging is partial or human-oriented
  • no structured action ledger
  • missing decision rationale capture

4.2 Silent regressions​

Symptoms:

  • metrics degrade slowly
  • correctness erodes without obvious breaking changes
  • error rates climb without clear root cause

Typical causes:

  • high change velocity
  • insufficient regression testing
  • lack of drift/anomaly detection

Category 5 — Security & Supply Chain Failures​

5.1 Dependency and toolchain risk​

Symptoms:

  • unsafe upgrades
  • pinned versions removed
  • compromised packages introduced

Typical causes:

  • agents “update everything” without risk scoring
  • lack of SBOM/lockfile protection policies
  • missing change classification gates

What ADG does about this​

ADG introduces:

  • enforced boundaries (tenancy, auth, environments)
  • verification gates (schema/RLS checks, doc sync, invariants)
  • a structured audit trail (action ledger + rationale)
  • drift detection and anomaly monitoring
  • escalation and human oversight at high-risk checkpoints