Agentic AI Coding for Earth Science

A 20-Minute Coding Task in 90 Seconds

The Unit of Work Has Changed

We are moving from writing code with AI to supervising AI that works on code.

By the End of This Lecture, You Will Be Able To

• Describe what a coding agent is, and what its loop does at each step
• Write a repository context file that states your scientific conventions
• Define a task as a contract: goal, context, constraints, evidence, stop
• Place a check at the right level of the verification pyramid
• Say which scientific decisions you may not delegate, and why

Today's Lecture

Last Week in Sixty Seconds

• Pseudocode before Python: get the logic wrong on paper, not in code
• A centred window and a trailing window answer different questions
• Three rungs of verification: it runs, it reproduces, it is defensible
• Today: hand the typing to an agent and keep every one of those rungs

Today's Repository

What Makes an AI Agentic?

Coding agent  =  Model + Context + Tools + Environment + Feedback loop + Permissions

The Scientific Agent Loop

A software agent loop stops at Observe. Verify and Record are what make it usable for science.

What Tools Can an Agent Use?

Three Products, One Mental Model

Prompt Engineering Is Not Enough

An Agent-Ready Task Is a Contract

Five words to carry out of this lecture: Goal, Context, Constraints, Evidence, Stop.

Teach the Repository, Not the Prompt

## The Water Year

The water year begins on October 1 and is named for the calendar year it
ends in. Water year 2018 runs 2017-10-01 to 2018-09-30.

Annual summaries in this project use the water year. On this record the two
conventions disagree by nearly a factor of two for 2018.

## Missing Observations

Missing discharge stays NaN. It is never filled with zero, and never
interpolated unless a human asks in writing and the choice is recorded.

Zero is a real discharge: it means the channel is dry.

An Earth Science Agent Needs More Than Coding Style

Passing Tests, Failing Science

$ python -m pytest tests -q
.....                                        [100%]
5 passed in 0.73 s

$ python -c "import streamflow; print(streamflow.annual_mean(df).loc[2018])"
4195.4   cfs     the record says the answer is 2418.0

Five green tests, and the number is 74 per cent too high. The suite checks shapes and exceptions. Nobody wrote a test for the year convention, because the person who wrote the tests had the same misunderstanding as the code.

Step 1: Read Before You Touch

Inspect this repository. Explain the analysis workflow and identify the
scientific assumptions it makes. Do not edit anything.

read  docs/SCIENTIFIC_CONTEXT.md
      water year begins October 1
      missing discharge stays NaN, never zero
read  src/streamflow.py
grep  'fillna|dt.year|m3/s'
      line 14  .fillna(0)           contradicts the missing-value rule
      line 20  .dt.year             calendar year, not water year
      line 38  'Discharge (m3/s)'   the values are cfs
      line 47  .dt.year again       the same decision, duplicated

Step 2: Separate the Two Kinds of Change

Step 3: One Change, Then Run It

- ax.set_ylabel("Discharge (m3/s)")
+ ax.set_ylabel("Discharge (cfs)")

$ python -m pytest tests -q

.....                                [100%]
5 passed in 1.24s

Step 4: The Science Got Better and the Test Went Red

+ def water_year(dates):
+     return dates.dt.year + (dates.dt.month >= 10)

- return df.groupby(df["date"].dt.year)[...].mean()
+ return df.groupby(water_year(df["date"]))[...].mean()
- df["discharge_cfs"] = df["discharge_cfs"].fillna(0)

$ python -m pytest tests -q

    def test_annual_mean_returns_one_value_per_year(df):
        result = streamflow.annual_mean(df)
>       assert len(result) == 3
E       assert 2 == 3

1 failed, 4 passed

Step 5: Put the Science Into the Test

def test_annual_mean_uses_water_years(df):
    # 2017-10-01 to 2019-09-30 is water years 2018 and 2019.
    # Three groups would mean it is grouping by calendar year.
    result = streamflow.annual_mean(df)
    assert list(result.index) == [2018, 2019]
    assert result.loc[2018] == pytest.approx(2418.0, abs=0.1)


def test_missing_days_are_skipped_not_zeroed(df):
    # Ten days are missing in June 2018. Filling them with zero
    # would pull the water-year mean down to about 2351.8 cfs.
    assert streamflow.annual_mean(df).loc[2018] > 2400

Step 6: What Did It Leave Behind?

$ python -c "print(annual_mean(df).round(1).to_dict())"
$ python -c "print(summary_table(df).set_index('year').mean_cfs.round(1).to_dict())"

annual_mean    {2018: 2418.0, 2019: 5764.3}
summary_table  {2017: 664.2, 2018: 4313.5, 2019: 5018.2}

6 passed

What the Agent Could and Could Not Know

The Scientific Verification Pyramid

Agents help at every level. Responsibility moves to you as you climb.

Earth Science Has Its Own Silent Errors

Create the Evidence Before You Delegate

reference_monthly = (stream.assign(month=stream["date"].dt.to_period("M"))
                     .groupby("month")["discharge_cfs"].mean())
reference_annual = (stream.assign(year=stream["date"].dt.year)
                    .groupby("year")["discharge_cfs"].mean())
print(reference_annual.loc[2018:2020].round(1).to_string())

year
2018    4263.5
2019    4712.1
2020    2775.8

A Tiny Test Is Manufactured Scientific Truth

tiny = pd.DataFrame({
    "date": pd.to_datetime(["2030-01-01", "2030-01-02", "2030-01-03", "2030-02-01"]),
    "discharge_cfs": [100.0, 300.0, np.nan, 50.0],
})
by_month = (tiny.assign(month=tiny["date"].dt.to_period("M"))
            .groupby("month")["discharge_cfs"])
print("mean: ", by_month.mean().tolist())
print("count:", by_month.count().tolist())
assert by_month.mean().loc["2030-01"] == 200.0

mean:  [200.0, 50.0]
count: [2, 1]

Beyond One Agent

When Parallel Agents Help, and When They Lie

• Genuinely useful: one agent on data checks, one on numerical tests, one on the record of what was done
• They are useful because the tasks are independent, not because there are three of them
• Five agents agreeing is not five pieces of evidence: they share a model, a prompt and a blind spot
• Agreement between an agent and a measurement is evidence. Agreement between agents is a correlation

From Prompt to Scientific Workflow

The agent owns 04 and 05. You own 01, 02, 06, 07 and 08. Steps 6 to 8 are in accent because they are the ones a software course leaves out.

What Should You Delegate?

The Policy, and an Honest Log Entry

Group Lab 4: Agentic Coding on a Broken Repository

The New Skill Is Supervision

• An agent observes, chooses a tool, and acts on the result: inspect, plan, act, run, observe, verify, record
• Put your conventions in the repository, not in a prompt
• Define the task as a contract: goal, context, constraints, evidence, stop
• A green suite is level 1 evidence; units, time and missing values live at level 3
• AI does not remove the need to understand code. It changes what that understanding is for