AutoPilot Drift

Steer Payload — inline patch of currency, region, and guard, released with signed audit.
Inject Workflow Drift — EMEA EUR vs USD flagged, $14,000 refund loop quarantined live.
Live flowcard — CRM Sync agent at 89.3% accuracy, streaming payloads in real time

Inspiration Autonomous agents now run CRM sync, outreach, and billing — and their most dangerous failure is silent. Not a crash, but silent workflow drift: an upstream field changes, a region returns a new shape, and a webhook starts sending euros where dollars were expected. There is no alert, no error, no crash — just a bad value quietly propagating until a $14,000 refund loop ships to customers. We built AutoPilot Drift so a human always holds the remote control when agent workflows drift.

What It Does Telemetry interception — Every payload is checked in real time against its live schema contract (v12) as it flows through the Enterprise CRM Sync & Outreach Agent queue, with 1,284 payloads streaming through a live event log. Anomaly flagging — When drift is detected, the workflow is quarantined before it writes: the [Assign Owner / Detect Drift] step flags amber, accuracy drops from 89.3% to 61.7%, and the console forecasts the blast radius (e.g. the EMEA EUR-vs-USD mismatch that would spin a $14,000 refund loop downstream). Live steering — The operator opens the inline [Execute Action / Steer] editor, patches the parameter remotely (currency code, region/routing, refund-loop guard, operator note), and releases the payload with one click — accuracy recovers to 94.2% and a signed audit entry is recorded. Pause bypass — With one tap the operator can pause the entire pipeline (queue held, nothing ships) and resume instantly — bypassing stale auto-flow whenever the situation needs a human decision. How We Built It A single-file, zero-backend HTML5 architecture: everything — the flowcard step tracker ([Check Request] → [Assign Owner / Detect Drift] → [Execute Action / Steer]), the anomaly detector, the steer editor, the queue controls, and the streaming log — lives in one index.html styled with Tailwind CDN (Plus Jakarta Sans + Newsreader serif, Salifact-inspired editorial layout on warm sand #F4F3EE with electric lime #D4F076). A vanilla JS state engine drives the drift/steer/pause state machine, accuracy bar, counters, toasts, and WebAudio alert beeps with no framework or build step. For the demo, an Edge-TTS voiceover (en-US-ChristopherNeural) was generated from DEMO_SCRIPT.md, and Playwright captured the live UI headlessly, with FFmpeg muxing the narration into the final demo_with_audio.mp4.

Challenges We Ran Into Simulating live drift schemas without heavy servers was the hardest part: we needed convincing, region-specific anomalies (EMEA currency mismatch, refund-loop blast radius) that felt like production telemetry but ran entirely client-side in a single file. We solved it with a lightweight inline state engine that mimics contract-checking, quarantine, and audit-signing semantics — no backend required. The second challenge was smooth visual transitions: making flagged/steered/paused states feel instant and trustworthy (bounded step pills, a dark animated success bar, slide-in anomaly cards, no layout shift or horizontal overflow on desktop or mobile) while keeping the page a single, self-contained HTML file that runs offline via file://.

Accomplishments That We're Proud Of Sub-20ms heuristic checks for payload validation in the demo engine, so drift is caught the moment it arrives. A clean editorial Framer-style UI (Salifact-inspired) that makes a technical safety console feel calm and human — centered hero, floating pill navbar, and a focused flowcard where one click injects drift, one click steers it, and one click pauses the world. Zero backend, double-click to run: judges can inject their first drift in ten seconds. And a fully reproducible demo pipeline — Edge-TTS narration, Playwright screen capture, FFmpeg muxing, all scripted and pushed to GitHub.

What We Learned That autonomy without remote control is a liability. The most valuable feature in the console isn't the automation — it's the pause button, the inline patch editor, and the signed audit trail. We also learned that a single-file demo can still feel like a real system if the state transitions (flagged → steered → safe) are legible at a glance, and that voiceover-driven video demos are dramatically more convincing when every click in the recording is timed to the script.

What's Next for AutoPilot Drift Production hardening: authenticated operator sessions, a real POST /check and POST /steer API behind the same inline contract checker, cross-agent drift correlation across multiple workflows, and notification routing (Slack / PagerDuty) when a pipeline is quarantined. Longer term: predictive drift forecasting from the payload history we already stream, so the console warns before the schema actually breaks.