Ship 30: locked topics + question bank
Umbrella (founder's words)
"How to make agents useful to all sorts of domains." The lean is enterprise agents, which trail cutting-edge work because they need governance controls and have to be made available inside trusted platforms. That is why Snowflake and Anthropic are a good vendor focus.
The 3 topics (LOCKED 2026-10-07; Ray proposed them, founder approved)
- Agents that actually work: the plumbing behind useful agents (the harness around the model, testing, quiet failures). Case study: Ray.
- Agents + data platforms: semantic layers, definitions in git, agent-ready data, governance inside trusted platforms (Snowflake, Anthropic). phData lane.
- Which jobs to give agents first, by domain: small business admin, healthcare billing, family logistics, kids' learning. [[agentify-3000-series|Fifth-job series]], Agentify 3000 scoring.
Question bank (first-hand, from a 2026-10-07 Claude Code activation session on the agentic software development lifecycle, ~30 advanced users)
Context: spec-driven development, where the agent writes intent.md, plan.md and test.md.
| # | Question | Founder's answer in the session | Ray's angle from our own build |
|---|---|---|---|
| Q1 | Are those documents thrown away? | Yes | Our dispatch briefs are single-use too. What survives is the decision log and the tests. |
| Q2 | Isn't there long-running context? How is it kept? How do you capture tacit knowledge? | A knowledge base. For enterprises, how to wrangle that context is still open. | Vault + working-context + memory files; corrections become rules. Enterprise version = governance (Topic 2). |
| Q3 | Is spec-driven development overkill? It makes sense for features, but bug fixes? | (open) | Scale the ceremony to the size of the change. A one-line brief plus a failing test is a spec. |
| Q4 | Agent code reviews catch things outside the spec. Do you triage them into a new task, send the work back to the builder, or add a test that catches it earlier? | (open) | Tonight's live example: a fresh critic found defects the brief never named, so they went back to the builder, each with a test, and the loop is capped (breaker). |
Rule: activation calls are a source of questions. Capture them first-hand and write about them. Do not persist client names or attendee details.
2026-10-08 addition: practitioner signal (LinkedIn DM, former Progress SQL analyst from the Mammoth engagement)
- Unprompted outreach after the cert post. He independently built a Cortex Agent with a semantic view + Cortex Search, keeps spec/BRD as markdown in the dbt repo, feeds it to Cortex Search, and builds a metric tree from existing reports "so the agent knows how things are calculated."
- Essay candidate (week 2, cite with his permission): "Which specs to throw away and which become your agent's memory." Plans are disposable (archive after ship, avoid context rot); definitions are durable (feed them to retrieval). Merges activation-call Q1 + Q2.
- Founder asked him (10-08 08:20) "what would you be interested in hearing?" → save his answer here.
2026-10-08 addition: AWS AI-DLC vs Claude AI-native SDLC playbook (founder-shared)
- Sources: https://aws.amazon.com/blogs/devops/ai-driven-development-life-cycle/ · https://claude.com/resources/articles/the-ai-native-sdlc-playbook
- AWS: 3 phases, "Mob Elaboration/Construction", "bolts"; light on verification, memory, governance. Claude: intent→spec→plan→PR→incident chain; CLAUDE.md + skills as memory; "the agent that wrote the code has no way to approve it"; commits as audit trail.
- Tension to resolve before posting Q1 (throwaway specs): playbook says "update plan.md in the same commit" (keep, audit trail) vs founder's "archive + gitignore after ship". Likely synthesis: plans archive; intent + decisions stay.
- Post idea: "Your spec isn't a document, it's an audit record."
2026-10-09 addition: activation call #4 (advanced Claude Code working session, 2026-10-08; founder's first-hand notes, no client details)
Pains heard, clustered:
- The laptop is the build server (5 of 11): parallel worktree agents choke local memory at compile time and overnight plans stall · CI runners getting longer (tests vs run time vs quality) · memory-hungry Playwright e2e locally and in CI · keeping the machine awake overnight · locked-screen office policy makes an unattended dev box a security problem. Their dodge of Claude Code cloud = couldn't configure it with GitLab. Session advice: run only the changed slice of the test suite.
- Context and maintenance: "agent oral history" (handoffs between agents; cached reads for subagents loading context) · skill maintenance ("I wrote this for Opus 4, still relevant?" → plugin eval CLI, ablation tests, check the memory too) · skills as text expanders (long reusable prompt behind a slash command) · Claude Code mods used to monitor builds in flight.
- Economics: Haiku x5 vs Opus one-shot. Cost-per-task vs benchmark accuracy = Pareto frontier; their pushback: that's one-shot, multi-turn may still be cheaper. Frame offered: the world splits into batch autonomous jobs and interactive jobs. Ray's angles from our own build: we hit nearly all of these on the Mac mini (4am restarts, permissions outage 09-20, fleet windows, handoff file = our oral history, the 07-25 skill retune for Opus 5 = skill maintenance). Post seeds: "Your laptop became a server and nobody told IT" · "Agent oral history: what one agent must hand the next" · "Price the solved task, not the token" (count retries and review turns).