Day 10: Principles for Building Agents — Three Levels of AI, the 7-Question Workflow-vs-Agent Test, Goldilocks Scope, Agent Prompting, and Tool Design

Level 1, the LLM: great at drafting, no access to anything external. Level 2, the AI-powered workflow: the LLM has tools (calendar search) but a human pre-defines every step and decision - the Day 8 Reel pipeline is exactly this, 'a fixed path' with AI at some steps. Level 3, the agent: built on a ReAct-style reason-then-act loop; the human supplies a goal ('build a statue that looks like this') and the model plans, selects tools and executes. Analogies stack up - Mario vs an RPG, railway vs taxi, donut machine vs chef, sculptor, gardener with a toolkit. The trap he names: adding AI to more and more steps does not promote a workflow to an agent.

Seven questions: Is the task complex or ambiguous? Do you know the destination but not the path? Does the value justify the resources? Are errors recoverable rather than catastrophic? Are the tools definable? Is there a clear step-by-step process? Would a workflow be too rigid? Four or more yes -> build an agent. Catastrophic-error domains - payroll, vendor payments, bank transactions - stay deterministic. The framework is meant to be screenshotted and pasted into an LLM (Gemini for reasoning) with the use case, returning a recommendation and confidence score.

Live scoring: 'voice AI agents for offices' -> agent (all yes); 'social media scheduling and posting' -> workflow. Rapid quiz: form validation = workflow; competitor pricing research = agent; customer-intent routing = AI-powered workflow; weekly calendar planning = agent; image resize/upload = workflow; debugging 500 errors = agent. 'You cannot wake up one day and say, I'm gonna build an agent.'

More tools raise the chance of confusion; more information raises token cost; more complex tasks lower reliability. The fix: the minimum information needed, a focused toolset, clear specific objectives - 'don't overwork your agents.' The same principle governs context ('too much context, agent will get confused. Too little, it might go off track') and tool-call budgets. Scale by specialization: one narrow agent per repeatable task (research, visualization, writing) coordinated by a manager agent - the Jerry pattern with its email, calendar and manager agents. Simple agents 'can solve big problems' and are easier to test, maintain and scale.

Chat prompting (role, context, task, format) prescribes the path - 'micromanagement.' Agent prompting is 'delegation' built on five pillars: the objective; the tools with explicit when-to-use guidance; a suggested approach; constraints (what to avoid, budgets - cap web searches); measurable success criteria that tell the agent when to STOP. Built live in Gemini by screenshotting the framework slide plus a plain description ('an AI data analyst with these tools', 300-word cap) for v1, then adding the 'components of an effective agent' slide (role, tools, input, behaviour, guardrails, success, error handling) for a ~500-word v2 with tool-selection guidance and stopping rules.

Failure modes to design against: the endless searcher ('keep searching until you find the highest quality source' loops forever - cap at 5-7); the answer-finder with no 'when you find it, stop'; unbudgeted expensive tools (5 calls for simple, 10-15 for complex); conflicting instructions ('be thorough, but also be quick'); vague success; 'what seems obvious to you may not be obvious for the agent.'

Clarity: action-oriented names ('web_search' not 'tool_a'), clear descriptions, unambiguous examples. Completeness: full parameters, comprehensive responses, error and edge-case handling. Uniqueness: one tool per job, no overlap - 'don't have a tool to do everything.' Chaining must be stated in the prompt (customer-lookup output feeds dispatch). Recap: naming, count, scope, documentation. Testing: 'don't break your head' on every edge case - 3-5 realistic tasks covering ~80% of scenarios, run manually, observe, iterate on real feedback. Personal agents can live on manual review; once an agent enters a business process (support), formal evals, metrics and monitoring apply.

The 90% rule: humans do the first ~5% (goal, constraints, context, success criteria, tool access, guardrails), the agent does the middle ~90% (tool use, research, multi-step work), humans review before shipping - Bolt building an app from a brief is the example. Then four real systems. (B) Bank transactions -> customer payments: pure workflow, ~2,000 transactions per run, filter debits, match customers, post payments, report - zero AI because errors are financial. (C) WhatsApp purchase orders -> Zoho sales orders: AI-powered workflow - acknowledge, branch on PDF vs text, AI parses 100+ line items, create the order, stock logic; ~30 minutes to ~1 minute. (D) WhatsApp hospitality agent: a true agent - check room availability and prices, generate a Razorpay payment link, update the CRM, message back, choosing tools by conversation state. (A) The live data-analyst agent: given get_data plus pie/line/bar QuickChart tools and the v2 prompt, 'split of MRR by product' produced a pie chart and 'revenue trending over time' a line chart, unprompted.

LLM vs MCP: the model thinks, the MCP is a capability it can call. Research outreach at scale: arXiv MCP + Apify scraping + an email workflow. Confidential clients: Pinecone or Google Vertex AI depending on ecosystem; local n8n with Ollama swapped for the OpenAI node for privacy; Ollama + LangChain for local RAG. Shiny-object syndrome: build one thing fully; validate with a waitlist landing page first; find ideas in your own leverage and network, not trends - lead generation is the universally sellable small-business automation. Harshad demos a 6-7 workflow master system: scrape prospects, verified-email filter, LinkedIn/company enrichment, hyper-personalized email, 3-step follow-up, 50 emails per inbox with rotation, reply detection halts the sequence - the $15-20k JSON is not shared. Claude Desktop custom connectors on Windows/free plan: enable developer mode and edit claude_desktop_config.json. Compare models by provider pricing pages or OpenRouter's unified key (GPT-5.2 at $1.75/M, 5.2 Pro at $21/M cited). Sighted: skills.sh (Vercel-linked skills aggregator), Mem0 / super memory for cross-session memory, Remotion for code-based video. Remote access to client PCs: integrate by API/JSON, never TeamViewer-style control. Closing: the course teaches 'the way to think,' not tools.