Skip to main content

​
DriftGuard

Some reliability problems don’t announce themselves. They accumulate — small behavioral shifts, growing memory deviations, patterns that move outside their baseline one session at a time. By the time they’re visible, they’ve been building for weeks. DriftGuard watches for them.

​
What Drift Looks Like

Drift is harder to catch than a crash because nothing breaks outright. Instead:
  • An agent that used to respond in 800ms now consistently takes 2.5 seconds — not slow enough to trigger a timeout alert, but degraded
  • Memory patterns shift over successive sessions — the agent’s context grows slightly faster each day without explanation
  • Behavior becomes less consistent: the same input produces noticeably different outputs across sessions, indicating something in the agent’s operating environment is changing
Standard monitoring misses all of this. DriftGuard doesn’t.

​
What DriftGuard Does

  • Establishes a behavioral baseline for each monitored agent
  • Tracks deviations from baseline across memory patterns, response characteristics, and context behavior
  • Flags long-horizon erosion before it crosses the threshold into active failure
  • Surfaces trends with the context to understand what’s changing and when it started

​
Signals DriftGuard Monitors

​
Context Drift

Changes in how the agent’s context behaves over time:
  • Context growth rate (tokens per session vs. baseline)
  • Compaction frequency and efficiency
  • Memory retention patterns across sessions

​
Baseline Deviation

Comparison of current agent behavior against established baseline:
  • Response latency distribution
  • Output consistency across similar inputs
  • Tool usage patterns
  • Error rate trends

​
Memory Integrity Posture

Whether the agent’s persistent memory state is remaining coherent:
  • Memory growth velocity
  • Anomalous retrieval patterns
  • State consistency across sessions

​
Quick Start

# Establish a baseline for an agent (run once)
acme driftguard baseline --agent my-agent-name

# Start drift monitoring
acme driftguard watch --agent my-agent-name

# View current drift status
acme driftguard status

# Compare current behavior to baseline
acme driftguard compare --agent my-agent-name

​
Output

DriftGuard Status
==================
Timestamp: 2026-03-09 15:42:31 UTC

Monitored Agents: 2

  pipeline-agent       📊 STABLE     Deviation: +2%  (within normal range)
  support-agent-prod   ⚠ DRIFTING   Deviation: +18%  ← ATTENTION

support-agent-prod — Drift Detail:
  Baseline established: 2026-02-20 (18 days ago)
  
  Context growth rate:  +23%  ⚠  (baseline: 420 tok/session → current: 516 tok/session)
  Response latency:     +12%  ↑  (baseline: 840ms → current: 941ms)
  Compaction frequency: +31%  ⚠  (was compacting 1.2×/session → now 1.6×/session)
  Output consistency:   -5%   ↓  (within range)

  Trend: 18-day upward trajectory. Started 2026-03-01.
  
  Recommendation:
    Review agent configuration changes from 2026-03-01 onwards.
    High compaction frequency suggests context growth outpacing budget.
    Consider increasing context budget or reviewing memory management.

​
Establishing Baselines

DriftGuard needs a stable operating period to establish a reliable baseline. When to baseline:
  • After a new agent deployment once it’s been running for a few days
  • After a major configuration change to reset expectations
  • Periodically (recommended: quarterly) to account for intentional evolution
# Establish baseline (captures current 7-day behavior window)
acme driftguard baseline --agent my-agent-name

# Baseline with custom window
acme driftguard baseline --agent my-agent-name --window 14d

# Reset baseline (use when you've intentionally changed agent behavior)
acme driftguard baseline --agent my-agent-name --reset

​
Drift Thresholds

DriftGuard uses configurable thresholds. Defaults are conservative — flag early, before problems compound. 
# ~/.acme/driftguard.yaml
thresholds:
  context_growth_rate: 15     # % deviation from baseline before warning
  response_latency: 20        # % deviation from baseline
  compaction_frequency: 25    # % deviation from baseline
  output_consistency: 15      # % drop from baseline before warning

monitoring:
  check_interval: 3600        # seconds between drift checks (default: 1h)
  trend_window: 7             # days of history for trend analysis

​
Integration with Sentinel and Agent911

DriftGuard operates on a longer time horizon than Sentinel:
  • Sentinel catches active failures (minutes)
  • DriftGuard catches slow degradation (days to weeks)
Both feed into Agent911. DriftGuard findings appear in the Agent911 snapshot when drift thresholds are exceeded.

​
CLI Reference

# Baseline management
acme driftguard baseline --agent <name>
acme driftguard baseline --agent <name> --reset
acme driftguard baseline list

# Status and comparison
acme driftguard status
acme driftguard status --agent <name>
acme driftguard compare --agent <name>

# Monitoring
acme driftguard watch --agent <name>
acme driftguard watch --all

# History
acme driftguard history --agent <name>
acme driftguard history --agent <name> --days 30

​
Next Steps

Sentinel

Active failure detection — the short-horizon complement to DriftGuard.

Agent911

DriftGuard findings surface in the Agent911 snapshot when thresholds are exceeded.