Field Notes - Jan 13, '26

Executive Signals

• Lock-in is the new outage: abstract models, not business logic; keep hot-swappable adapters and failover
• Events before databases: capture first, parse later to avoid blind spots
• Attachments are acquisitions: treat portals, auth, and checksums as core data
• Agents loop, budgets don’t : hard limits prevent spend spikes and side effects
• Org keys beat shadow creds: RBAC and logs make failures visible, reversible

Engineering

Provider-Agnostic Code Agent Stack

Recent ToS crackdowns and capacity throttles proved “unlimited” isn’t unlimited. Don’t couple workflows to a single LLM or bespoke CLI. Abstract the model boundary, not business logic, so you can swap providers without rewriting core systems.

• Define a thin interface; keep adapters for two providers hot and config-flippable
• Enforce org-owned keys with RBAC and centralized logging
• Auto-failover if rate-limit/5xx >1% for 10 minutes or median latency doubles

Event-First Email Ingestion, Not DB Polling

Polling vendor-managed databases adds lag, masks failures, and drops attachments. Treat inbound mail as events: capture losslessly first, then parse. Make object storage your source of truth; any vendor DB is only a convenience mirror.

• Webhook → durable queue → worker writes raw RFC822 to object storage; only metadata to DB
• Store attachments as objects with content hashes; target ≥99.9% capture and <60s ingest-to-available
• Own retention, restores, and quotas; treat vendor DB as a mirror, not authority

Attachments As Data Acquisition

Compliance artifacts often arrive as PDFs or portal-only downloads requiring auth, sessions, or JavaScript. If you only parse email bodies or MIME parts, you’ll miss the artifact you’re measured on. Build for both modalities with provenance.

• Build an authenticated fetcher for “download” links; snapshot HTML/PDF with checksum and provenance
• Track expected vs captured by source; alert if misses >2% for 15 minutes
• Provide a fallback: notify owner with deep link and SLA; log manual retrieval for audit

Guardrails For Autonomous Agent Loops

Give agents tools and they’ll loop forever. Prevent runaway spend and unintended writes with hard budgets, timeouts, and human checkpoints. Detect repetition early and stop automatically with clear traces for debugging.

• Enforce per-run token/credit budgets and wall-clock timeouts; hard stop on either
• Require human checkpoints before external writes or high-cost calls
• Detect repetitive tool calls (>N) and auto-halt with a trace dump