| name | anti-drift |
| description | Prevents agent divergence from task intent through checkpoint validation, scope monitoring, and alignment gates. Detects scope creep, time overruns, and error cascades, then forces re-alignment before work continues. Use when: multi-agent tasks, complex implementations, long-running sessions. Skip when: trivial single-step tasks, pure research.
|
| tags | ["orchestration","quality","agents"] |
| triggers | ["check for scope creep","agent drifting off task","are we still on track","validate task alignment","campaign work going off course"] |
Anti-Drift — Agent Alignment & Divergence Prevention
Purpose: Agents solving complex tasks will drift from the original intent. This skill
detects drift early and forces re-alignment before wasted work compounds.
Why This Matters
Without anti-drift:
- Agent starts fixing a bug, ends up refactoring 12 files nobody asked for
- Multi-agent task has each agent making conflicting assumptions
- 30-minute task balloons to 2 hours because scope expanded silently
- Agent retries the same failing approach 5 times instead of escalating
With anti-drift:
- Checkpoint every N steps forces "am I still on track?"
- Scope monitor flags when files touched exceed the plan
- Error cascade detector stops thrashing after 2 consecutive failures
- Time overrun detector flags when work exceeds 2x the estimate
The Three Drift Detectors
1. Scope Creep Detector
Trigger: Agent touches files not in the original plan or routing decision.
How it works:
- At task start, the routing decision lists expected files/directories
- After each edit, compare touched files against the plan
- If files touched exceed plan by >3 files → SCOPE CREEP ALERT
Response to scope creep:
DRIFT DETECTED: Scope Creep
Planned files: [list from routing decision]
Actual files touched: [current list]
Excess: [N files outside plan]
OPTIONS:
A) Revert to plan — undo out-of-scope changes
B) Expand plan — update routing decision to include new files (requires justification)
C) Escalate to CC — ask if scope expansion is warranted
Config: .agents/config.toml [anti_drift.signals] scope_creep_threshold = 3
2. Error Cascade Detector
Trigger: 2+ consecutive errors/failures in the same task.
How it works:
- Track each execution step's success/failure
- If 2 consecutive steps fail → ERROR CASCADE ALERT
- This prevents the "keep trying the same thing" anti-pattern
Response to error cascade:
DRIFT DETECTED: Error Cascade
Failed step 1: [description + error]
Failed step 2: [description + error]
MANDATORY ACTIONS:
1. STOP executing. Do not attempt fix #3.
2. Switch to alternative approach (BRAIN_LOOP Step 4 alternatives)
3. If no alternatives exist, generate them NOW
4. If 3 total approaches have failed → ESCALATE TO CC
Config: .agents/config.toml [anti_drift.signals] error_cascade_limit = 2
3. Time Overrun Detector
Trigger: Task is taking >2x the estimated time.
How it works:
- At routing, estimate task duration based on complexity tier:
- TRIVIAL: 2 min | SIMPLE: 10 min | MODERATE: 30 min | COMPLEX: 60 min | ARCHITECTURAL: 120 min
- Track elapsed time from task start
- If elapsed > 2x estimate → TIME OVERRUN ALERT
Response to time overrun:
DRIFT DETECTED: Time Overrun
Estimated: [X minutes]
Elapsed: [Y minutes] ([Z]x overrun)
Current step: [what's happening now]
CHECKPOINT:
1. Is the original approach still viable?
2. Should we simplify scope?
3. Is there a blocking issue that needs CC's input?
4. Should we commit partial progress and reassess?
Config: .agents/config.toml [anti_drift.signals] time_overrun_factor = 2.0
Checkpoint Protocol
Every N task steps (default: 5, configurable in config.toml), the executing agent must pause and answer:
CHECKPOINT [step N]:
Original intent: [1-sentence task summary]
Current state: [what's been done so far]
Remaining: [what's left to do]
On track: [YES / DRIFTING — explain]
Files touched: [count] / [planned count]
Errors encountered: [count]
If "DRIFTING" → apply the relevant drift detector response above.
Validation Gates
Gates are hard stops between phases. Work cannot proceed until the gate passes.
Pre-Execute Gate
Before any code is written:
Post-Execute Gate
After implementation, before commit:
Pre-Commit Gate
Before any git commit:
Post-Deploy Gate
After pushing to remote / Vercel deploy:
Multi-Agent Coordination Rules
When multiple agents work on the same task (COMPLEX+):
- One coordinator — The architect agent owns the overall plan
- Isolated scopes — Each agent works on assigned files only
- Sequential handoffs — Agent B waits for Agent A to finish before starting dependent work
- Shared memory — All agents read/write to the same ACTIVE_TASKS.md
- Conflict detection — If two agents modify the same file, flag immediately
- Max concurrent agents: 4 (configurable in config.toml [anti_drift.max_concurrent_agents])
Integration Points
- BRAIN_LOOP Step 6 (EXECUTE): Checkpoint protocol runs during execution
- BRAIN_LOOP Step 10 (HEAL): Drift metrics logged to MISTAKES.md if drift occurred
- Task Routing Skill: Routing decision provides the baseline for scope monitoring
- Config: All thresholds and limits in
.agents/config.toml [anti_drift]
Obsidian Links
- [[brain/BRAIN_LOOP]] | [[brain/AGENTS]] | [[skills/task-routing/SKILL]] | [[brain/CAPABILITIES]]
- [[memory/MISTAKES]] | [[memory/PATTERNS]]
Maven-specific adaptation
Marketing work is high-risk for drift because every campaign has dozens of "while we're at it..." adjacent improvements (rebrand fonts, refactor tracking, redo landing page). Maven's rule: ship the smallest creative iteration that proves the thesis, then capture follow-ups via agent_inbox to itself or Bravo. Don't refactor the funnel while shipping a single ad set.
