AI in Production

When a version bump lands in a PR, a GitHub Action triggers Claude to read the diff and draft a structured changelog entry — categorized by type (new, improved, fix), with human-readable descriptions.

Manual changelogs are tedious and often skipped. AI-generated drafts maintain consistency and reduce the friction of shipping transparently. The human still reviews and edits, but the 80% grunt work is gone.

Users submit feedback in natural language. Claude (Haiku) analyzes the message, cross-references it against recent changelog entries and existing open feedback, and produces a structured ticket with type, priority, title, and description. Duplicates are flagged automatically.

Raw user feedback is noisy — bug reports mixed with feature requests mixed with confusion. Automated triage turns unstructured input into actionable tickets without a PM manually reading every message.

The agent can create its own triggers: time-based ("check back on this market research in 3 days"), condition-based ("alert me when task completion drops below 70%"), and inactivity-based ("nudge if the user hasn't logged in for a week"). A background worker evaluates triggers every ~60 seconds.

Most AI tools are purely reactive — they wait for you to type. A proactive agent that sets its own reminders and follow-ups behaves more like a real co-founder. This is one of the hardest patterns to get right because it requires persistent state, reliable scheduling, and contextual intelligence.

The agent runs async research sessions over the live web. Each session can span multiple iterations — searching, reading, synthesizing, then searching again with refined queries. Results are stored as structured artifacts (findings, competitors, market data) that persist in the workspace.

Single-query web search gives shallow results. Multi-iteration research with refinement mimics how a human analyst works: start broad, identify gaps, drill deeper. Storing results as workspace artifacts means the research compounds — future decisions can reference past findings.

A background process generates personalized daily email digests for each user. The brief summarizes: what happened yesterday, what decisions are pending, what tasks are due, and what the AI recommends focusing on today. Delivery timing is configurable per company.

Email is where attention lives. A well-timed daily brief reduces the need to open the app to stay informed. It also creates a natural touch-point that drives engagement without being pushy — the brief is useful, not spammy.

Every decision the AI surfaces is classified into one of three tiers. Operational decisions (low impact, reversible) the AI handles autonomously. Tactical decisions (moderate impact) it presents with a recommendation for human approval. Strategic decisions (high impact, hard to reverse) it blocks until explicitly approved. The AI can never resolve its own decisions.

Trust is the #1 barrier to AI adoption in business. Most tools either give the AI too much power (scary) or too little (useless). A tiered governance system earns trust incrementally — you see the AI make good operational calls, so you trust its tactical recommendations, and eventually its strategic advice carries weight.

A single user message can trigger a chain of 10+ tool calls. Example: "Research our competitors and update the plan" might execute: fetch business context → run web search → analyze results → update research artifacts → read current plan → modify plan sections → create follow-up tasks → log activity. All in one turn, with the LLM deciding each step.

The difference between a chatbot and an agent is execution. Chatbots give advice. Agents do the work. The orchestrator pattern — reason, act, observe, repeat — lets the AI perform complex multi-step workflows that would take a human 30 minutes in under 60 seconds.