clone & run:
github.com/lusiki/AI_workshop_wiiw
verify prices at claude.com/pricing

AI in Economics Research — the working sheet

From frontier ideas to a working agentic pipeline · a seminar for wiiw

next task Describeyou: request Planagent: read-only Approveyou: plan gate Executeunder permissionsagent: allowlist Verifyagent + critic Commityou: review diff

The five components — & the failure each prevents

#ComponentPrevents
1Context — CLAUDE.md + path-scoped rulesagent reverting to generic defaults
2Plan gate — read-only plan, approve oncesolving the wrong problem fast
3Permissions — command allowlist, denied pathsan unbounded autonomous process
4Verification — render/test, critic→fixer, numeric provenanceconfident nonsense in the paper
5Git — small commits, quality score, worktreesrepeated & unrecoverable mistakes

Binds together: context = competent · plan gate = aligned · permissions = contained · verification = honest · git = reversible.

What it costs (USD / 1M tokens)

TierModelinout
CheapHaiku 4.515
MidSonnet 5315
HighOpus 4.8525
FrontierFable 51050

Batch −50% · cache reads ≈0.1× input · route 70/20/10 (Haiku/Sonnet/Opus). Subs: Pro $20/mo · Max $100 & $200 · Team ≈$25/$125 seat.

Data governance

  1. Restricted microdata never enter a prompt — denied paths; the model reads only aggregated outputs.
  2. Prefer Team / Enterprise / API; check training-data terms before onboarding.
  3. EU-region endpoints for residency; pin the tool version in shared config.

Reading list · A tool · B instrument · C object

  1. Korinek 2023 JEL — taxonomy A
  2. Korinek 2025 NBER — chatbots→agents A
  3. Dell 2025 JEL — DL as data creation A
  4. Bahoo et al. 2025 — adoption map A
  5. Sant'Anna 2026 — the practitioner manual
  6. Ludwig, Mullainathan & Rambachan 2025 — two valid uses B
  7. Athey, Keleher & Spiess 2025 — nudge whom B
  8. Gavrilova et al. 2025 — forests in DiD B
  9. Chen, Cheng, Liu & Tang 2026 — theory-guided TL B
  10. Brynjolfsson, Li & Raymond 2025 QJE — field evidence
  11. Chatterji et al. 2025 — usage at scale
  12. Acemoglu 2025 — sober benchmark C
  13. Jones 2026 — weak links vs acceleration C
  14. Brynjolfsson, Korinek & Agrawal 2026 — the agenda C
  15. Mullainathan 2025 — objects & instruments
1. Three lenses — know the flagship paper for each. 2. Five parts — context, plan gate, permissions, verification, git. Start there. 3. The binding constraint is your verification, not the machine's generation.
wiiw · The Vienna Institute for International Economic Studies Print to PDF: landscape A4 · one page