ReAct Agent Pattern

Curriculum teaching stub for ReAct loops — explicit thought → action → observation with bounded iterations. Used in VAP Deep Research.

▶ Live demo · Architecture · Portfolio · Pattern series

What this is

Part 1 of 5 in Curriculum Agent Patterns — tool-using assistants, retrieval workflows, and enterprise Q&A with inspectable traces.

How we solve it

Problem	Approach
Monolithic LLM calls	Explicit ReAct loop with stop conditions
Runaway tool loops	Bounded iterations + schema-light validation
Untestable agents	Deterministic model stub — pytest without API keys

Case study & tradeoffs

No standalone case study — see venkat-ai.com/work and VAP case study. Tradeoffs in docs/ARCHITECTURE.md.

---

Scope: Curriculum stub with deterministic tests and a live trace viewer — not a production agent fleet. Compose into Venkat AI Platform for governed graphs.

Status

Component	Status	Notes
Pattern demo + trace UI	✅	Scenario chips · simulated metrics · compose CTA (VAP/OmniForge)
Core agent loop	✅	Reference implementation
LangGraph production graph	🟡	Teaching scope — compose into VAP for fleet use
MCP tool bridge	❌	See LoopForge / VAP MCP docs
AegisAI gateway	❌	No side effects in pattern demo
Pytest regression	✅	pytest -q in repo

Agent skills (Cursor + Codex)

Org skills: vpeetla-ai-skills.

git clone https://github.com/vpeetla-ai/vpeetla-ai-skills.git
./vpeetla-ai-skills/scripts/install.sh --cursor --codex --project .

---

Quick start

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
python -m react_agent_pattern
pytest

The default demo uses a deterministic model stub so the architecture can be tested without external API keys.

cp .env.example .env

For LLM keys, database config, and production adapters, see docs/LOCAL_DEVELOPMENT.md.

Repo layout

src/react_agent_pattern/
  __main__.py       # CLI demo
  agent.py          # ReAct orchestration loop
  models.py         # Model protocol + deterministic demo model
  tools.py          # Tool interface and sample tools
  tracing.py        # Structured trace events
docs/
  ARCHITECTURE.md   # Architect-level design decision record
tests/
  test_react_agent.py

Path to production (compose into VAP)

Replace ScriptedReasoningModel with an LLM gateway that supports tool-call structured output. Keep boundaries intact: the agent should not know vendor SDK details, and tools should remain independently testable.

Interview map

Business function: Minimal LangGraph teaching stub for one agent pattern (compose into VAP for production).

Staff+ prep crosswalk — playbook · study UI · Practice Arena · org matrix. Only entries this repo honestly exercises.

Category	Entry	Fit
System design	Agent tool-use / orchestration (md)	Pattern slice only — not a full platform
Coding	Clone a graph (cycle-safe) (md)	Light — graph structure intuition for LangGraph

Related

• Next in series: Reflection Agent Pattern
• Composed pipeline (not this stub): AI Content Factory — multi-agent content orchestration with HITL

If this helped you, ⭐ the repo — and star the series to follow new patterns.