Episode 0: TLDR: one request, a team of agents

AI for Research Efficiency, episode 0, with Douglas Hutchings. Narration: an AI-generated voice (ElevenLabs).

The fastest way into real work with a command-line AI agent. Do a few quick steps, then ask in plain words: say who you are, what you want at the end, and "spin up a team of agents". In the run this episode replays (recorded on 22 October 2026, Claude Code 2.1.308 in its default auto mode), one request took an empty folder to a result in 12 minutes 31 seconds: the lead agent set up the project, three agents worked at once on floods, low flows and the river's whole pattern, and the team finished with five figures and a one-page summary of 86 years of the Buffalo River's records.

Your run will look different. Models update, and every computer is set up its own way. If something goes wrong, ask your agent to help fix it: paste the error and say what you were trying to do.

The quick part (Windows first, the Mac beside it)

1. Make a folder for your project, right-click inside it, and choose Open in Terminal (Windows 11, File Explorer; on some PCs it is under Show more options). Mac: Control-click the folder in Finder's path bar, then Open in Terminal.

2. Install one agent, with one line. Choose one; each has its own episode.

   Agent	Windows (PowerShell)	Mac (Terminal)	Episode
   Claude Code	irm https://claude.ai/install.ps1 | iex	curl -fsSL https://claude.ai/install.sh | bash	2.1
   Codex	powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"	curl -fsSL https://chatgpt.com/codex/install.sh | sh	2.2
   Antigravity	irm https://antigravity.google/cli/install.ps1 | iex	curl -fsSL https://antigravity.google/cli/install.sh | bash	2.3

   Lines read from each vendor's page on 22 October 2026; check the vendor's page if one fails.

3. Open the terminal again, the same way (so it finds the new program), and type the agent's name: claude (or codex, or agy).

4. Sign in the first time, with a monthly plan or an API key (your browser opens), and trust the folder when it asks: in Claude Code the choice starts on "No, exit", so press Down, then Enter, for "Yes, I trust this folder".

The request

The run's request, word for word. Put in your own field, data and result:

I'm a hydrologist studying the Buffalo River in Arkansas. Spin up a team of agents and take this empty folder all theway to an interesting result. Set it up so you remember how the project works and I can go back to earlier versions,pull the long-term USGS gauge records, and find out what has changed in how the river floods and runs low. Finish withfigures for a paper and a one-page summary.

Its shape: I'm a ___. Spin up a team of agents and ___. Finish with ___. Examples to try (not run for this episode):

• I'm a historian of the Ozarks. Spin up a team of agents and turn this folder of scanned letters into a searchable archive. Finish with a timeline of who wrote to whom.
• I'm a soil scientist. Spin up a team of agents and clean up this year's instrument exports into one dataset. Finish with a quality report for each run.
• I'm a statistics instructor. Spin up a team of agents and build a practice dataset for each week of my course. Finish with a worksheet and an answer key.

What the team did (the recorded run)

• The lead agent (0:00 to 2:08): made a Python environment, found the Buffalo River's long-term USGS gauges, downloaded their daily and peak records, wrote a shared loader "so every agent parses the files the same way", wrote its own instruction file (CLAUDE.md), and saved a first version in Git (tagged v0.1-data).
• Three agents at once (2:08 to 10:13): floods, low flows, and the flow regime with the rain. The low-flow agent reported first ("a solid null result"), then the regime agent ("Flow rose about 16%, and wetter precipitation accounts for nearly all of it"), then the flood agent ("no detectable change in flood magnitude, frequency, or timing").
• The lead again (10:13 to 12:31): drew five figures and checked each, said "I'll check a few numbers I want to quote first", rebuilt everything from the raw data ("every table comes out identical"), wrote the captions and the one-page summary, and tagged v0.2-analysis and v1.0-paper.

The result, in the summary's own title: "A wetter river, not a changed one". Mean flow at the St. Joe gauge was about 16% higher in 1983 to 2025 than in 1940 to 1982 (checked by the studio against USGS's daily data: 1,012 and 1,173 cubic feet per second). By the team's analysis, more rain explains nearly all of it, and floods and low flows show no detectable change. The team listed its own caveats (rain for the whole climate division stands in for the basin's; where the record is split; low flows near the gauge's limits); read them before you rely on a number.

Before you use your own data

Watch episode 9, Your Data, Your Keys, Your Campus: what an agent sees, where keys belong, and your campus's rules. And when you want to know why each step works: episode 2.0 (the terminal), 2.1 to 2.3 (the tools), 4 (the instruction file and everyday habits), 5 (versions and sharing), 8 (many agents at once).

Sources

• Install lines: Anthropic, Set up Claude Code; OpenAI, the Codex README and Codex CLI; Google, Antigravity Get started; read 22 October 2026.
• Auto mode: Anthropic, Choose a permission mode (auto mode is the built-in starting mode from Claude Code 2.1.283; "Auto mode reduces permission prompts but does not guarantee safety."), read 22 October 2026.
• CLAUDE.md: Anthropic, How Claude remembers your project, read 22 October 2026.
• Data: U.S. Geological Survey, daily and annual peak streamflow for 07056000, Buffalo River near St. Joe, AR (water years 1940 to 2025), retrieved 22 October 2026; the team also used NOAA's nClimDiv precipitation for Arkansas division 1.