Side Quest: Deterministic vs Agentic Data Ops

August 26, 2026 · View on GitHub

Optional: use this guide when you are unsure which parts of a data workflow should stay deterministic and which parts should be agentic, then return to Step 16.

Data workflows work best when you split jobs on purpose. Keep repeatable operations deterministic. Use the agent when you need judgment.

:clipboard: Before You Start

The decision rule

Use this quick test:

  • If you can define exact pass/fail logic in advance, keep it deterministic.
  • If you need to decide which of 40 open issues is most urgent, make it agentic.
  • If you need to explain trend changes to leadership, make it agentic.

You do not need one mode for the whole workflow. Most production workflows are hybrid.

Data-ops examples

TaskBetter fitWhy
Fetch the last 24 hours of commitsDeterministicSame command, same shape, every run
Count open P1 incidents from issue labelsDeterministicExact filter and count logic
Decide which incidents look most urgent to humansAgenticNeeds contextual judgment
Summarize trend changes for leadershipAgenticRequires interpretation and audience-aware writing
Validate JSON schema before downstream useDeterministicFixed validation rules
Explain likely causes behind a change spikeAgenticHypothesis and narrative reasoning

Hybrid blueprint

Follow this structure for repository status, incident triage, and reporting flows:

  1. Deterministic extraction: run fixed commands (gh, git, API calls) to collect data.
  2. Deterministic shaping: normalize and label outputs ($GITHUB_OUTPUT, JSON fields, counts).
  3. Agentic interpretation: ask the agent to identify risk, priority, and notable patterns.
  4. Agentic communication: ask for role-specific output (engineering digest, leadership summary, on-call handoff).

This keeps your pipeline reliable. It also gives you flexible reasoning where scripts become brittle.

:hammer_and_wrench: Try it: Label each step D or A

Read the workflow snippet. In the comment block, label each step as D (deterministic) or A (agentic).

# Step A: Fetch open issues from the last 24 hours.
gh issue list --state open --search "updated:>=2026-07-13" --json number,title,labels,updatedAt

# Step B: Shape the output into a sorted table with issue number, label count, and last update time.

# Step C: Decide which three issues need maintainer attention today and explain why.

# Your labels:
# Step A: _
# Step B: _
# Step C: _
Show answer key
  • Step A: D — fixed command and fixed fields.
  • Step B: D — fixed transform and sort rules.
  • Step C: A — requires prioritization and explanation.

Common anti-patterns

  • Pushing raw noisy logs straight into the prompt without shaping them first
  • Asking the agent to compute exact metrics that shell commands can compute reliably
  • Hard-coding dozens of branching rules for narrative tasks that change every week
  • Using the agent for safety checks that require strict deterministic guarantees

:white_check_mark: Checkpoint

  • You can explain the difference between deterministic and agentic work in one sentence
  • You can identify one step in your workflow that should stay deterministic
  • You can identify one step in your workflow that should become agentic
  • You can describe a hybrid design for your current data workflow
  • You know when deterministic validation should remain outside the agent

Return to Connect a Live Data Source to Your Workflow.