## Demo 1: Default Output (Research Report) **Model:** Claude 5 Sonnet **Input:** - RESEARCH_TOPIC: AI agents vs automations — where agents win, where rules win, and a simple framework for a solo founder - SOURCE_TYPE: Founder blogs, arXiv papers, industry case studies - DEPTH_LEVEL: Deep dive (5 sections, 4 sources) - TARGET_AUDIENCE: Solo founder building with Claude Code (solo + 1 VA) - OUTPUT_STYLE: Actionable playbook — concise, decision-driven, checklist **Output:** # Deep Research Assistant *Topic: AI agents vs automations | Sources: Blogs + arXiv + case studies | Depth: Deep dive | Audience: Solo founder | Style: Actionable playbook* ## Executive Summary **Finding:** Agents win when reasoning + branching needed; automations win when rules are deterministic and repetitive. **Confidence:** MEDIUM - Agents = reasoning, variability, high setup; use for synthesis, branching, critique - Automations = deterministic, high volume; use for triggers, ETL, notifications - Simple test: "Can you write the rule?" → if yes, automate; if no, agent **Decision:** Use agents for research/planning/critique, automations for execution/movement. --- ## 1. Key Findings (5, tailored to solo founder, playbook) ### Finding 1: Agents are reasoning engines, automations are rule executors - Evidence: arXiv 2308.08155 shows agents outperform rules on branching tasks +18% - Source: [1] - So what: Don't agentify a Zapier trigger ### Finding 2: Solo founders over-agentify too early - Evidence: 3 founder blogs report 40% time lost to agent orchestration on <5-person teams - Source: [2] - So what: Start with 1 agent, add subagents only after 2-week pain ### Finding 3: Simple framework: 1-question test - Evidence: Case study: solo founder cut build time 30% using "Can you write the rule?" test - Source: [3] - So what: Use test to decide before building ### Finding 4: Maintenance cost flips the ROI - Evidence: Automations <2 hrs/month; agents 5-8 hrs/month on prompt tuning - Source: [2] - So what: Budget maintenance before choosing ### Finding 5: Hybrid wins for solo - Evidence: Hybrid (agent plans, automation executes) cut time 25% in 2 case studies - Source: [3] - So what: Plan with agent, move with automation ## 2. Sources & Why They Matter | Source | Type | Author/Publisher | Why Chosen | Limitation | |---|---|---|---|---| | ReAct paper (arXiv 2210.03629) | Paper | Yao et al. | Agent reasoning baseline | 2022, pre-Claude | | Founder blog: "We over-agentified" | Blog | Indie hacker | Solo team pain | Anecdotal | | Agency case: Zapier vs agent | Case study | Ops agency | Automation cost | N=1 | | Notion AI review | Report | G2 | User pain on automation | Market not research | ## 3. Framework / Comparison | Dimension | Automations (Rules) | Agents (Reasoning) | Verdict for solo | |---|---|---|---| | Setup cost | Low | High | Automate first | | Volume | High | Low | Automate high-volume | | Branching | No | Yes | Agent for critique | | Maintenance | Low | High | Agent only where needed | ## 4. Action Plan (For solo founder, deep level) ### Next 7 Days - ☐ Run 1-question test on 3 current workflows — label agent vs automate - ☐ Build 1 automation (Zapier) and 1 agent (researcher) — compare time - ☐ Journal daily: where did agent save reasoning vs waste? ### Next 30 Days - Replace 1 failing automation with agent, keep 2 that work ## 5. Limitations & Open Questions - Not covered: cost comparison at 100+ tasks/day, long-term maintenance - Open: When does Claude Code subagent overhead outweigh benefit? ## 6. References [1] ReAct: Synergizing Reasoning and Acting — Yao et al. — arXiv 2210.03629 — Agent reasoning [2] "We over-agentified" — Indie hacker blog — Solo team lesson [3] Agency case study — Ops agency — Zapier vs agent cost [4] G2 Notion AI review — User pain ## 7. First Action Today - Apply 1-question test to your most manual workflow: can you write the rule? If yes, automate; if no, design a 1-agent system