Slides

# Workflow patterns · optional reference

## Further reading after Week 6

These patterns extend the feedback-loop theory discussed in Week 6. The
humanifier case study and graphical-tree lab illustrate feedback at different
scales. This page is further reading.

---

# Two questions to keep separate

## How much work is handed over—and how does the workflow run?

**The eight levels**

- How much help the AI provides
- Which context and tools it receives
- How independently it can continue
- How much review is required

</div>
<div>

**The workflow design**

- What repeats in a loop
- Which paths are fixed in code
- Where branches or parallel work occur
- Which conditions stop or escalate work

</div>
</div>

A small loop can be sufficient; branches or additional agents are useful when
the task requires them.

<p class="source-note">Eight-level framework adapted from Bassim Eledath, <a href="https://www.bassimeledath.com/blog/levels-of-agentic-engineering">The 8 Levels of Agentic Engineering</a>.</p>

---

# The agentic loop

## Gather → act → verify → repeat until a stop condition

<div class="workflow-row">
  <div class="workflow-step"><strong>Gather</strong><br>task, state, evidence</div>
  <div class="workflow-arrow">→</div>
  <div class="workflow-step"><strong>Act</strong><br>respond or call a tool</div>
  <div class="workflow-arrow">→</div>
  <div class="workflow-step"><strong>Verify</strong><br>schema, test, metric, review</div>
  <div class="workflow-arrow">↺</div>
</div>

- A weak verifier lets the loop produce increasingly polished mistakes
- A deterministic stop condition may be a passing test, bounded score, turn cap, or human handoff
- Domain review covers what software tests cannot: does the result answer the intended life-science question?

<p class="source-note">Architecture guidance: <a href="https://www.anthropic.com/engineering/building-effective-agents">Anthropic, Building effective agents</a></p>

---

# Workflow or agent?

## Who determines the path through the system?

**Workflow · code directs the path**

- Known sequence and branches
- Predictable latency and cost
- Easier testing and audit
- Prefer for stable scientific procedures

</div>
<div>

**Agent · model directs the path**

- Steps are not known in advance
- Flexible tool selection
- Higher variance and wider evaluation burden
- Reserve for tasks where flexibility earns its cost

</div>
</div>

Many reliable applications combine a fixed outer workflow with one bounded agentic step.

---

# Five reusable workflow patterns

## Add orchestration only when the simpler path is insufficient

<div class="course-grid">
  <div class="course-card"><strong>A → B → C · Chaining</strong><p>Fixed stages with validation gates between them.</p></div>
  <div class="course-card"><strong>Route · Specialize</strong><p>Classify first, then select a prompt, model, or tool path.</p></div>
  <div class="course-card"><strong>Parallel · Aggregate</strong><p>Split independent work or compare multiple attempts.</p></div>
  <div class="course-card"><strong>Lead → workers</strong><p>Let an orchestrator create subtasks and synthesize results.</p></div>
  <div class="course-card"><strong>Generate ↔ evaluate</strong><p>Iterate when criteria are explicit and feedback is actionable.</p></div>
</div>

The simplest adequate architecture may contain no agent at all.

<p class="source-note">Patterns adapted from <a href="https://www.anthropic.com/engineering/building-effective-agents">Schluntz & Zhang, Building effective agents</a>.</p>

---

# Course navigation

[Week 6: feedback loops](week_06.html) · [Week 7: agentic RAG](week_07.html) · [Course map](course_map.html)