Time: 20–30 min · one sitting · no code.
This is the Thursday (2026-08-06) orientation from the field-map
research — vocabulary + compass, not a multi-day pack.
Working definition
Agentic Software Engineering (ASE): tool-using language-model loops are the primary implementers and maintainers inside sandboxed interfaces and machine-checkable gates. Humans own goals, product taste, and risk.
Not autocomplete tips. Not “I use coding agents” as proof you can ship product agents.
Two lanes (don’t collapse them)
| Lane | Meaning | Proof looks like |
|---|---|---|
| L1 · Coding agents | Agents write/maintain your software | Tickets closed, plan gates, repo harnesses |
| L2 · Product agents | Agents are the product | Tools, state, evals, services users depend on |
Career-target roles need L2 proof. Daily study is L2-weighted on purpose.
Eight competence axes (field map)
- Briefing — machine-executable tickets; two agents same pass/fail
- Harness / tools / MCP — bounded tool I/O; sandboxes
- Context / knowledge — repo maps, structured retrieval (not only vectors)
- Verify / evals — gates + behavior goldens
- Failure / safety — max turns, blocklists, durable stop
- Human authority — taste, release, blast-radius decisions
- Economics — $/task, routing, token budgets
- Org design — review for spec alignment, not only syntax
Failure taxonomy → tag every golden
When you write or score a golden, name which failure it catches:
| Failure | Control idea |
|---|---|
| Amnesic spec drift | Re-inject the brief each turn |
| Hallucinated API / stubs | Schema / world checks on tool use |
| Infinite loop inflation | Max steps + duplicate-call halt |
| Test gaming | Don’t let the agent edit goldens/tests |
| Blast-radius escalation | Sandbox + refuse dangerous tools |
| Silent architectural regression | Side-effect + policy checks; human taste on structure |
What changes for study — and what doesn’t
| Question | Answer |
|---|---|
| Change P0 this month? | No. Still eval harness → Python agent service. |
| Why not GraphRAG-first? | External 30/60/90 often leads context/maps first. Useful backlog — wrong first hole when product evals are still the empty skill. |
| What does change? | ASE vocabulary, failure tags on goldens, clear parked list (MCP / GraphRAG / durable multi-agent later). |
| Day-to-day task after this page? | eval-harness-101 Day 1 |
Pack queue (current)
- orientation-ase — this page (once)
- eval-harness-101 — active P0
- python-agent-service-101 — next
- product-loop-101 — after that
- Parked: context-maps, mcp-tools, durable-multiagent, failure-guardrails
Self-check
- ASE in one sentence?
- Name the two lanes — which is the usual job-target gap?
- Why not start with GraphRAG this month?
- List three failure modes you’ll use when writing goldens.
Log idea: pack orientation-ase ·
did: field-map skim + self-check · one-line takeaway.