AmpCode's Thorsten Ball on how to use Agents to connect build, release, and production observability

How I Work with Agents: Letting Agents Connect Build, Release, and Production Observability Horizontally

@thorstenball (@AmpCode) has distilled "how to use Agents" into an executable worldview: software is getting cheaper, rate of change must enter planning, the "pile up slop" argument is outdated, prompts should state what not how; what truly amplifies returns is the triangle of Agent × codebase readability × process, and letting Agents horizontally cover build → release → CI/production logs/feature flags/monitoring, not just replacing typing in the editor.

How he sees the present · The time constant for fixing bugs, changing features, and producing prototypes has changed; things previously considered troublesome like feature flags and Grafana are now worth handing to Agents. · You can't define processes only by "what models can do in October 2026"; you must factor in token price drops and frontier updates roughly every two weeks—we may be in a deflationary period where "six months from now it's cheaper and faster." · He believes "not reviewing every line will pile up into unmaintainable slop" no longer holds: there will be slop, but strong models can also help you clean it; last year models were often confused, now you can throw very dirty code at them to fix. · Prompting goes up a level: no longer emphasizing file @ mentions and complex formatting; e.g., "double the speed of the bytecode compiler and VM," "build a similar project based on these screenshots." Amp Web/macOS/iOS have removed @ mentioning.

Triangle: Agent · Codebase · Process Just stuffing Agents into "the old text editor slot" and then pushing/reviewing as before won't raise GDP. Let Agents vertically master frontend/backend/libraries/infrastructure, then horizontally deploy, watch CI, read production logs, flip flags, and check metrics. Amp's own narrative: a monorepo of ~2 million+ lines, almost entirely built by AI; after adopting Orbs/cloud Agents, commit velocity noticeably increased.

Personal daily usage · Always use the strongest model (names Opus 5.5); cheap/open-source models suit fixed workloads, otherwise multi-round correction may cost more. · Dictation is heating up; no longer manually managing context (handoff, thread map, etc. are de-emphasized). · Agents don't write prose for humans (Slack is hand-written); they write mechanical copy: troubleshooting steps, incident report parameter descriptions, etc. · Automations/scheduled wake-ups become essential: check error rates two hours after release; use "snooze until next message" in the sidebar to keep it clean. · Don't wait idly for Agents: advance asynchronously; when changing UI, request multiple variants at once and pick, rather than iterating back and forth. · Only review load-bearing code; edges (like CSS) have higher fault tolerance. Mantra (from Quinn): "Give me irrefutable evidence that this works."—E2E, screenshots, screen recordings, test matrices; with proof in hand before review, the mindset is completely different. · For about 8 weeks he barely wrote code locally; occasional chores on the local machine / running amp --no-tui on a Pi for home automation.

Team: Making the codebase "readable" to Agents · About 15 AGENTS.md + skills (dev server, gcloud, BigQuery, deployment, etc.): being able to see, start, and inject data reduces hallucinations. · Threads as shared memory: Agents can read_thread to trace a colleague's discussion on a certain commit. · All-in Orbs: one temporary machine per thread, scaled to zero when idle; install dependencies/FFmpeg/build slideshows freely. When macOS-specific needs arise, use Runner; Orb and local threads exchange files and proof. · Production side: canary, logs, production paths dedicated for Agents (analogous to "changing code for tests"); plugins to flip flags, check health, read bug diagnostic bundles. · Process leans AI-native: push to main, few/no PRs (avoid Orb running 20 minutes of CI then re-running remotely), designers also prototype and fix code in Orbs, no Figma. · Click a bug in Slack → Agent pulls diagnostics → fixes → sends email/grants credits; incidents use get system health-type plugins. · Evaluating Agent-friendliness: after changing AGENTS.md, spawn multiple clean Orbs (including a control group) to verify whether the new scripts/plugins are actually used, then merge.

Case: Orbs OOM Report received in Slack → start an Opus thread to check top memory processes and alerts → check logs/BigQuery/similar bugs → iterate on the plan → wrap in a feature flag → release → human approves enabling the flag → spawn another Orb to artificially consume memory for verification → author sends back a screen recording for confirmation. Emphasis: small-step releases to production + flags are far cheaper with Agents than the old local clicking around.