Agent-oriented software engineering

lastmod 2026-04-17created 2026-03-022 edits

What is agentic coding?

Say "build a sign-up feature" and something does come out. The problem is that it is likely not what you wanted
Experience of handing the AddPlan screen to an agent with just a PRD, then losing half a day to dozens of turns of back-and-forth
A 5-minute Socratic interview with the AI → edge cases sorted out in advance → revisions cut to 2-3 turns
Taking time to think before implementing — those 5 minutes save 4 hours

Writing AGENTS.md well matters, but if the code architecture itself is well designed, the speed at which the agent grasps the context is completely different
The agent was lost in a flat directory; restructuring into per-Feature directories brought immediate improvement
Armin Ronacher: "Tools need to be prepared for the LLM chaos monkey using them completely wrong"

Left a CLI project running overnight and it finished in 1 hour — only the type definitions were set up, and the business logic was an empty shell
On the second attempt, the agent rewrote the tests themselves to suit itself
The agent's "done" is different from my "done"
Elvis's 7-step DoD system (PR→CI→3 code reviews→Telegram) is extreme but shows the direction

In a redistribution engine, the same parameter had different semantics in each function → an infinite loop where fixing A breaks B
Retrying with the same prompt is banging your head against the wall in the same direction over and over
Classifying failures into 3 types (lack of context, wrong direction, structural conflict) makes the remedy clear
A "Must NOT Have" guardrail broke the infinite loop

Handing liquidglass to an agent and thinking "that's odd... let it be" was the most expensive decision
20 files tangled together, impossible to roll back
Afterwards, a tracer bullet strategy + blueprint — you cannot draw a blueprint in advance for a technology you are applying for the first time, so tracer bullets sketch it quickly
Observability builds trust, and trust makes delegation possible

Working 3 days in a row, you lose 15 minutes every morning explaining the context
Automatic memory extraction at session end with Claude Code hooks → 5-second restore in the next session
Boris Cherny's team checks CLAUDE.md into git so the whole team shares it
A structure where the team's memory, not an individual's, is passed to the agent

Boris Cherny runs 10-15 parallel sessions at once
The experience of managing 6 squads as a CTO is surprisingly similar to managing agents in parallel
Not ADHD but intentional multitasking = managing
People ask questions, but agents proceed on their own judgment without asking — up-front design matters more

Level 0 (coding directly) → Level 1 (instructing agents) → Level 2 (orchestrator) → Level 3 (meta design)
Experience of turning a 20-minute daily routine into a skill and cutting it to 2 minutes
Compounding engineering — a project is not a finish-line game but a compounding game. Earlier sessions compound into later ones

AI-made designs score 60-70 points. The moment Ellie's design comes in, there's a sense of "ah, this works"
An AI-made informational post got 0 likes, while an impulsive one-line brag post got 30,000 views
KinglyCrow's "No Skill, No Taste" — LLMs lowered the skill barrier to entry, but the real barrier, taste, is if anything amplified
Chris Lattner: "The more implementation is automated, the more design, judgment, and taste matter"
In an era flooded with the 80%, differentiation comes from the remaining 20%

author Zer0 sitemap