agentliveness

Autonomous AI systems fail silently — fresh output over a dead engine. Predict less, detect more.

A small, dependency-free reliability harness for scheduled / autonomous agent systems. It makes the failure mode that birthed it impossible to ship silently: state that looks alive but resets every run.

The failure it prevents

Under a scheduler (launchd, cron, k8s CronJob) every run is a fresh process. Any counter or accumulator held only in memory — initialized in __init__, never restored from disk — silently resets to its starting value on every run. The code reads correctly. In-process tests pass (one long-lived interpreter hides the bug). In production it is dead, and nothing tells you.

This is real: a network monitor's adaptive-cadence counter was always 0 at decision time because each scheduled process started fresh, so the adaptive behaviour never engaged — invisible for weeks behind green tests and a fresh timestamp.

Phase 1 — PersistentState

Restart-safe state, durable by construction:

from agentliveness import PersistentState

st = PersistentState("~/.myengine-state.json", default={"runs": 0})
data = st.load()
data["runs"] += 1
st.save(data)        # atomic; survives crash and process death

• Atomic writes (tmp + os.replace) — a crash mid-save never leaves a torn file that loads as garbage but looks fine.
• Versioned envelope + generated_at — schema drift fails loud; staleness is checkable.
• Load-or-default — a missing or corrupt file recovers as "first run" instead of crashing the engine.

The test that is the thesis

tests/test_restart.py does not re-instantiate an object in one interpreter — it spawns real subprocesses that increment, save, and exit, proving the counter accumulates 0→1→2→3 across genuine process death. A test that runs in-process would pass even with the bug; this one reproduces the production execution model.

pytest

Phase 2 — LivenessContract

Freshness asks "was this written recently?" Liveness asks "is the thing that writes it actually working?" — strictly stronger. A fresh file over a dead producer passes a freshness check and fails a liveness contract.

from agentliveness import LivenessContract

c = LivenessContract(
    path="~/.myengine.json",
    max_age_s=2 * 3600,
    producing=lambda payload: bool(payload.get("norms")),  # the 'actually working' signal
)
v = c.evaluate()
if not v.healthy:
    alert(v.reason)        # e.g. "fresh but EMPTY — producer emits nothing"

Bundles four invariants — exists · fresh · non-empty · producing — into one verdict. A freshness-only monitor calls a fresh-but-empty file healthy; this catches it. Warmup-honest: an unwarmed subsystem reports warming, not degraded, so first boot does not cry wolf. Reads a PersistentState envelope or bare JSON — the two primitives compose.

Phase 3 — LoudFail

Detection without a response channel is an incident no one sees. LoudFail routes a verdict to sinks (log / macOS notification / exit code) — but only on a state transition, so a scheduled check that is still-down stays silent instead of training you to mute it. And a sink that throws (a broken notifier) is swallowed: it can never crash the run it is protecting.

from agentliveness import LoudFail, log_sink, notify_sink

lf = LoudFail("my-agent", "~/.my-agent-loudfail.json",
              sinks=[log_sink(), notify_sink()])
lf.report(contract.evaluate())   # fires once on down-edge + once on recovery; never raises

Incident state is persisted (via PersistentState), so "new incident" is judged across scheduled processes — the three primitives compose into one harness.

Phase 4 — agentliveness audit

Catch the bug before it ships. A static scan (stdlib ast only) that flags the never-restored-accumulator class in any agent repo: an attribute seeded as an empty counter / list / dict / set in __init__, grown across calls, but never restored from disk — alive in a long-lived process, silently reset every run under a scheduler.

agentliveness audit path/to/agent           # scan a file or directory
agentliveness audit . --json                # machine-readable (for an empirical scan)
agentliveness audit . --exit-zero           # report only; don't fail CI

⚠ engine.py:3  AdaptiveMonitor.consecutive_quiet  (counter)
  self.consecutive_quiet is seeded as an empty counter in __init__ and
  accumulated (line 6), but never restored from persistence. Under a scheduler
  every run is a fresh process, so it resets to its seed value on every run.

Exit code is non-zero when findings exist, so it gates CI. It's advisory — each finding is a candidate to either restore (PersistentState) or confirm intentionally per-run. It is also a Python API: from agentliveness import audit_path.

Roadmap

• Phase 1: PersistentState + the subprocess restart test. ✅
• Phase 2: LivenessContract — producing, not just fresh. ✅
• Phase 3: LoudFail — fire once per incident, never crash the run. ✅
• Phase 4: agentliveness audit — static scan for the failure class in any repo. ✅

MIT licensed.