Live run: a looping agent is stopped mid-flight — decision signed, audited, resumable.

Agents fail quietly. They retry the same broken test 40 times, burn $8 in tokens on a loop, run rm -rf on a database directory, and write to files they were never supposed to touch. The agent doesn't know it's failing — from its perspective, it's still trying.

Edward sits between the agent and its runtime. It watches the event stream, builds a picture of what the agent is actually doing across turns, and intervenes when the picture stops looking right.

Agent (Pi / Codex / custom)
    │ events
    ▼
Canonical Event Schema   ← normalizes tool names to capabilities
    │
    ▼
State Engine             ← materializes cross-turn agent state
    │
    ▼
Trigger Rules [FROZEN]   ← deterministic safety + convergence checks
    │
    ├─ HARD_CONSTRAINT ──→ Edward: BLOCK (scorer cannot override)
    │
    └─ SOFT_DECISION ──→ Local scorer ──→ Policy Resolver
                              │                │
                              └────────────────┘
                                       │
                                       ▼
                                 Control Kernel
                                       │
                                       ▼
                              PAUSE / CANCEL / RESUME

Why not just if/else?

A watchdog (error_count > 5 → stop) looks at individual events. The failure modes that actually kill long-running agents don't show up in any single event — they emerge from the shape of the trajectory over time. We tested this directly on held-out data (seed 137, frozen config):

Watchdogs miss temporal failure modes entirely because they have no memory across turns. Edward maintains a sliding window of tool calls, tracks file modifications, and computes elapsed time — none of which fit in a single-event check.

Measured, not claimed

Edward validates itself against StepShield (NeurIPS 2026), the first benchmark treating intervention timing as a first-class metric (9,429 trajectories, step-level labels, 216 held-out):

The deterministic layer alone is quantitatively blind to content-semantic violations (7.4%) — the "silent corruption" gap — while keeping the best false-positive rate. Adding a local 4B scorer with evidence-grounded task-contract probes and asymmetric temporal confirmation lands in LLMJudge-tier timing territory at zero marginal cost. Full measurement series and reproduction commands: BENCHMARK.md.

What it detects

Eight trigger rules, tuned on a dev split (seed 42) and frozen for held-out evaluation (seed 137):

Policy packs make the knobs yours: conservative / balanced (= FROZEN defaults) / aggressive, as TOML or JSON.

Quickstart

pipx install edward-guard            # zero dependencies, Python 3.11+

edward doctor                        # environment checks
edward demo                          # self-running proof: 6 failure scenarios, PASS/FAIL

edward wrap -- pi "fix the flaky test"                     # full monitoring + intervention
edward wrap --no-scorer -- python my_agent.py              # any command, rule-only
edward wrap --scope ./src --auto-resume 60 -- pi "task"    # scoped writes, auto-resume

Interventions are resumable, not fatal: PAUSE exits with code 75, pins the agent session, and edward wrap --continue picks the same session back up from the audit log. CANCEL / BLOCK exit 76. Audit lands in ~/.edward/audit.jsonl — including an estimated avoided-spend per intervention.

The scorer is optional and always advisory. Point EDWARD_SCORER_URL at any local OpenAI-compatible scoring endpoint (a 4B model on your GPU box is plenty — see deploy/ for the team-LAN topology). Scorer down? Edward logs a warning and runs rule-only. It stays protective.

v0.2.0 highlights

• Signed evidence receipts — every audit record is Ed25519-signed into a hash chain (pure stdlib, RFC 8032 vectors); edward verify proves tamper-evidence offline. Publish your public key; anyone can check.
• Human approval loop — --wait-approval 300 sends Resume/Kill links to Slack (or stderr) and waits; PAUSE becomes a decision, not a dead end.

Why zero dependencies?

Edward's control loop runs stdlib-only: it must boot on any Python 3.11+ box, inside any container, in front of any agent — including air-gapped ones. The heavy lifting (scoring) is delegated to a separate local service, which you own and can swap (4B quantized, bigger, whatever) without touching the control plane.

Repository map

edward/                the package
  cli.py               wrap / demo / eval / audit / doctor
  engine.py            ControlPlane: events → triggers → scorer → decision → audit
  state_engine.py      cross-turn agent state
  triggers.py          8 rules, policy-parameterized (defaults FROZEN)
  scorer_client.py     /v1/score client + circuit breaker
  stepshield.py        external benchmark adapter (EIR metrics)
  scenarios.py         failure scenario suite (demo/eval source of truth)
benchmark.py           300-trial held-out benchmark
robustness_eval.py     4-dimension robustness attack
BENCHMARK.md           full measurement series + reproduction commands
deploy/                team-LAN deployment templates

Status & roadmap

• v0.1.1 on PyPI, CI on three platforms
• StepShield integration with paper-aligned EIR metrics
• Robustness suite as edward eval --suite robustness
• Scorer fine-tune (targets FPR; data flywheel from audit logs)
• Cloud fleet console (team tier)

Contributing

Deterministic layer stays deterministic: trigger defaults are FROZEN, and behavior-affecting changes require re-running the benchmark gate. See CONTRIBUTING.md.

License

MIT — © 2026 Edward contributors