Error Budget (Agentic)

SRE error budget concept adapted to autonomous agent loops. The budget defines how much failure — in retries, cost, runtime, or error rate — is tolerable bef…

Error Budget (Agentic)

SRE error budget concept adapted to autonomous agent loops. The budget defines how much failure — in retries, cost, runtime, or error rate — is tolerable before the system must stop, narrow its capabilities, or roll back. Without a budget, self-healing loops run indefinitely.

---

SRE Origin

Error budget (GitLab/Google SRE definition): the acceptable amount of unreliability over a time window, derived from the Service Level Objective (SLO).

error_budget = 1 - SLO_target

Example: SLO = 99.9% → error_budget = 0.1% of requests may fail per month

When budget is exhausted: freeze feature releases until reliability recovers. The budget is a shared contract between engineering and the business on how much downtime is acceptable.

---

Agentic Adaptation

In agent loops, “unreliability” maps to multiple resource axes simultaneously:

Budget type	What it measures	Exhaustion action
Retry budget	Attempts per failure type	Stop retrying → rollback
Token budget	LLM tokens consumed per task	Escalate → don’t continue spending
Runtime budget	Wall-clock time per session	Timeout → checkpoint → stop
Error rate budget	Failure rate over rolling window	Rollback staging → halt deploy
Session budget	Total autonomous iterations	Human review before continuing

An agent loop without budget definitions can spend unbounded resources on a task it cannot solve, accumulating debt (bad code, flaky fixes, wasted compute) while reporting “still working.”

---

Budget Policy Rules

Rule 1: Define budget before starting, not during failure Budget values must be set in config, not chosen adaptively by the agent. An agent under pressure will rationalize higher budgets.

Rule 2: Budget exhaustion stops, not pauses When any budget is exhausted: stop, rollback if applicable, write incident note. Do not “just try one more time.”

Rule 3: No progress = budget consumption If the agent makes N attempts with identical failure signatures and zero test improvement, that is budget exhausted regardless of remaining retry count. Signature-based detection catches this without counting.

Rule 4: Token/time exhaustion escalates, not loops If the agent consumes its token or runtime budget without completing: checkpoint state, stop cleanly, flag for human. Looping on a budget-exhausted task is the failure mode, not the fix.

---

Reference Budget Defaults (Starting Point)

Calibrate these per project; these are conservative starting values:

budgets:
  retry_per_failure: 3          # max attempts before rollback per distinct failure
  retry_total_per_session: 10   # total retry attempts across all failures in one session
  token_per_task: 50000         # LLM tokens per discrete task (not per session)
  runtime_per_task: 1800        # seconds (30 min) per task; wall-clock
  staging_error_rate: 0.01      # 1% of synthetic checks may fail before rollback triggered
  session_iterations: 20        # total ralph-loop iterations before requiring human check-in

---

Budget as Rollback Trigger

Error budget integrates with Self-Healing Loop rollback conditions:

staging deploy → health monitor starts
  every 30s: check /health + run 1 synthetic test
  error_rate = failures / total_checks

if error_rate > staging_error_rate budget:
    rollback to last known-good image
    write incident note
    set ESCALATE=1
    stop

if agent retry_count > retry_per_failure budget:
    same action

The budget threshold replaces “wait and see” with a machine-enforceable condition. The health monitor runs independently of the agent — agent cannot suppress or override it.

---

Detecting No-Progress Without Budget

When an agent loops on the same failure with different patches but no improvement, retry count alone doesn’t catch it:

progress_score = (passing_tests_now - passing_tests_before_repair) / total_tests

if progress_score <= 0 and retry_count >= 2:
    # Agent is not making progress — escalate even if budget remains
    trigger_rollback()

This is the “spending without improving” detector: progress score of 0 across two retries = budget exhausted regardless of counter.

---

Budget vs Retry Count

These are related but distinct:

Concept	Scope	What it prevents
Retry count	Per failure event	Infinite loops on one broken thing
Error budget	Across the session/deployment	Accumulated failure across many events
Progress score	Per repair attempt	Retrying useless fixes

All three are needed. Retry count alone doesn’t stop an agent that tries 100 different fixes on 100 different things, each failing once. Budget covers the aggregate; progress score covers the per-attempt signal.

---

Related Pages

• Self-Healing Loop — the retry/rollback/escalation pattern budgets govern
• Agentic CI/CD — where error rate budgets are monitored (staging health)
• Ralph Loop — the outer loop that session budgets constrain
• Verification Pipeline — what passes/fails that feeds the error rate