Field Notes - Oct 31, '25

Executive Signals

Agents are the new interns: give autonomy in bursts with approvals and review
Heavy for autonomy, light for flow: route long codepaths to deep models, keep pairing fast
Cache is a feature: design explicit cached states, rescan paths, and instant clarity
GDPR-first by default: region selection upfront prevents rework and false reassurance
Parallel beats perfect: run multiple long agents and keep the strongest branch
Simpler CRM, faster truth: standard objects answer GTM without brittle add-ons

A reusable library of 3-5 second mascot clips outperforms a single three-minute hero video for websites and paid. Build a tiny internal generator to render character plus text into motion graphics and ship packs weekly. Fix lip sync and trigger logic once in the generator so every export benefits.

Deliver 10 reusable clips before revisiting long-form video
Centralize timing and lip sync fixes in the generator
Use clips to animate empty states, banners, and ad variants

Sales

Cut HubSpot Complexity; Ship a Simple Lead→Deal Flow

Avoid five-figure add-ons and fragile custom objects until the basics are proven. Standard contacts and deals, lifecycle stages, and CTA click tracking answer most go-to-market questions with less breakage and faster reporting.

Map MQL → SQL → Opportunity in one pipeline, delete unused custom objects
Instrument the top three CTAs end-to-end from page to meeting or opportunity
Re-import with a clean property schema and document side effects

Product

Region-First Compliance UX

If a public audit exposes regulatory risk, the region selector is not optional. Default to the strictest regime and make the choice explicit up front. It prevents false reassurance, clarifies scope across design and copy, and helps Sales by logging selections for later conversations.

Ship a mandatory region selector defaulting to GDPR
Preload 2-3 presets with plain-English differences
Log selections for analytics and downstream sales use

Ephemeral Reports, Persistent Value

For open scanners, treat shared reports as temporary artifacts. A 48-hour expiring link reduces retention risk and cost while preserving value with a timestamped PDF download, a clean Rescan path, and a visible cache banner.

Expire public links at 48 hours and offer timestamped PDFs
Add a Rescan button and label cached responses with scan date
Document that historical diffs require saved projects

Design for Cached Scans, Not Just Fresh Runs

Most UIs only design fresh and loading states. Real usage needs four: fresh scan, in-progress, cached result, and error. Cached should be instant and explicit. Error needs a one-click recover path and a short diagnostic code.

Implement all four states and instrument time saved from cache
Surface what changed only when persistence exists, otherwise prompt rescan
Treat errors as first-class with retry and a diagnostic code

Engineering

Choose Copilot Mode by Blast Radius

Use the least-privileged mode that still moves work. Ask will not touch files, ideal for legacy or fragile repos. Edit is a bridge to proposals. Agent can modify code and run terminal commands with approvals. Treat Agent like a powerful junior paired engineer.

Start in Ask for unknown codepaths; graduate to Agent when you can explain the diff
Keep terminal approvals on; require checkpoints before schema or codegen changes
Timebox Agent bursts to 10-20 minutes, review, then continue

Heavy Models for Autonomy, Light Models for Flow

For long, multi-file tasks, use heavyweight agents that think deeper and navigate big repos for 60-90 minutes hands-off. For interactive pairing and quick iterations, step down to lighter, faster models. Expect slower first tokens from heavy models, paid back by fewer re-prompts.

Route refactors or features over 30 minutes to heavy agents and batch context
Keep Q&A and small edits on lighter models to preserve developer flow
Budget for slower first token and reclaim time with fewer re-prompts

Reset Threads When Context Fatigue Hits

As threads grow, models drift near context limits. Latency spikes and answers degrade. Switching models mid-thread is fine, but windows differ. Checkpoint the state, then restart from a crisp brief to restore quality.

Dump current state and plan to a markdown brief, then start fresh
Cap threads at about 100 turns or after major dependency changes
If switching models, re-seed with one concise message of requirements and constraints

Parallelize Long-Running Agents to Cut Tail Risk

For hour-plus runs, spawn multiple attempts in parallel and choose the best trajectory. It is cheaper than discovering a single long run took a bad path. Select for tests passing, minimal surface area, and convention alignment.

Use four parallel runs for tasks over 60 minutes and compare diffs
Kill weak runs early based on failing tests, excessive churn, or style drift
Pick the branch that passes tests, minimizes changes, and matches conventions

Guardrails for Automating Third-Party Submissions

External portals often auto-notify on submit. Treat adapters like production-impacting code. Default to dry-run and require explicit human escalation before any real submission. Keep secrets isolated per environment and produce reviewable artifacts.

Ship a dry-run flag and block outbound submits until two-key approval
Isolate and rotate secrets per environment, never embed credentials in code
Emit would-send payloads or screenshots for review before first live run

Observe Then Harden for Bots

Do not over-engineer bot protection on day one of a public tool. Ship, observe real abuse patterns, then add minimal friction where it matters. Protect the rescan endpoint first since it is easiest to abuse and costliest to serve.

Log IP and device, alert at concrete thresholds
Add per-IP throttles and simple puzzles only after abuse crosses a set rate
Harden rescan endpoints before broader protections