Data-Analysis Agent Pattern: reliable data analytics

Pattern Essence

Data-Analysis Agent is a pattern where the agent works with tabular or time-series data through a controlled analytics pipeline: policy check, ingest, profile, transform, analyze, validate, report.

When to use it: when you need to compute and verify metrics on real data, not describe them in generic words.

Unlike "just answer", the agent performs step-by-step analytics:

• Ingest + profile: reads data and checks quality
• Transform: cleans outliers and missing values by rules
• Analyze: computes metrics, KPI, and aggregates
• Validate: checks result consistency
• Report: builds conclusions and artifacts for audit

Problem

Imagine you ask:

"How much did we earn last week?"

You get a number, but you do not understand what happened to data before calculation.

Critical questions remain:

• were refunds included
• were duplicates removed
• how missing values were handled
• whether data parts were dropped during "cleaning"

Analytics without a fixed process gives numbers that are hard to explain, verify, and reproduce.

That is why different people can calculate the same dataset differently.

That is the core problem: without a transparent pipeline, analysis output becomes unpredictable and non-auditable.

Solution

Data-Analysis Agent works through a fixed analytics pipeline.

Analogy: this is like lab analysis. First, they take a sample and check quality, then they compute indicators. Without this sequence, numbers cannot be trusted.

Core principle: value is not only in a number, but in the fact that each analysis step is explainable and reproducible.

The agent cannot "just calculate immediately" - it must pass the steps:

1. Policy check: verify access to source
2. Ingest: load data
3. Profile: inspect structure and quality
4. Transform: clean by rules
5. Analyze: compute metrics
6. Validate: verify result

And it must record:

• data source
• cleaning rules
• executed checks
• data limitations

This makes it possible to show, for example:

• which rows were removed
• how missing values were handled
• which outliers were detected
• which invariants passed

Works well if:

• ingest step cannot be skipped
• cleaning rules are fixed
• validation step is mandatory
• execution layer does not allow bypassing pipeline steps

Reliable analytics is a pipeline the agent cannot bypass technically.

How It Works

This pattern often uses Code-Execution Agent to run calculations in a sandbox.

Key difference: the focus here is not code execution itself, but a full analytics cycle with quality control gates.

In Code, It Looks Like This

decision = policy_engine.evaluate_source(source, user_role=user_role)
if decision.type != "allow":
    return stop_with_reason("source_policy_denied")

df = load_data(source)
profile = profile_data(df)

if profile.schema_mismatch:
    return stop_with_reason("schema_mismatch")

df_clean = clean_data(df, rules=cleaning_rules)
metrics = compute_metrics(df_clean)

checks = validate_metrics(metrics, rules=[
    "no_negative_revenue",
    "conversion_between_0_1",
])

if not checks.ok:
    return escalate_or_fallback(checks.errors)

report = build_report(metrics, checks, profile)
return report

Analytics output must include not only numbers, but evidence: which checks passed and what data limitations existed.

How It Looks During Runtime

Goal: prepare weekly sales summary

Ingest:
- sales.csv for last 7 days

Profile:
- 2% missing in channel
- 1 outlier in revenue

Transform:
- fill channel = "unknown"
- remove duplicate id=8472

Analyze:
- revenue: 142,300
- conversion: 3.84%
- top channel: paid_search

Validate:
- all invariants passed

Report:
- KPI table + short conclusions

When It Fits - And When It Doesn't

Good Fit

Not a Fit

Because Data-Analysis agent needs extra engineering: pipeline, checks, monitoring, and source versioning.

How It Differs From Code-Execution

Code-Execution is an execution mechanism. Data-Analysis is a process for getting reliable analytical output.

When To Use Data-Analysis (vs Other Patterns)

Use Data-Analysis when you need to explore data and return conclusions based on analysis.

Quick test:

• if you need to "explain what is happening in data and provide conclusions" -> Data-Analysis
• if you need to "just run code and get technical output" -> Code-Execution Agent

How To Combine With Other Patterns

• Data-Analysis + Code-Execution: when you need to compute or transform data, the agent runs code in sandbox with limits.
• Data-Analysis + Guarded-Policy: if data is sensitive, policies limit which tables and fields can be read.
• Data-Analysis + Fallback-Recovery: if source or step fails, pipeline switches to a safe recovery path.

In Short

Pros and Cons

FAQ

Q: Can an agent do analysis without data profiling?
A: It can, but that is risky. Without a profile step, schema problems are easy to miss and metrics can break.

Q: Why validate metrics after calculation?
A: To catch anomalies before sending output: negative revenue, conversion outside [0,1], and other invariants.

Q: Does Data-Analysis agent replace BI system?
A: No. It complements BI: automates one-off analysis, cleaning, and explanation, but does not replace data governance.

What Next

Data-Analysis agent works with structured data and reproducible metrics.

How do you build an agent for open-world research: source discovery, reading, fact extraction, and citation?