48 lines
18 KiB
JavaScript
48 lines
18 KiB
JavaScript
import{c as e,Q as n,j as t,m as s}from"./chunks/framework.BPKcPtvA.js";const u=JSON.parse('{"title":"Non-Technical Track: Decision Frameworks for AI Agents","description":"","frontmatter":{},"headers":[],"relativePath":"modules/non-technical.md","filePath":"modules/non-technical.md","lastUpdated":1780488246000}'),i={name:"modules/non-technical.md"};function o(r,a,l,d,p,c){return n(),t("div",null,[...a[0]||(a[0]=[s(`<h1 id="non-technical-track-decision-frameworks-for-ai-agents" tabindex="-1">Non-Technical Track: Decision Frameworks for AI Agents <a class="header-anchor" href="#non-technical-track-decision-frameworks-for-ai-agents" aria-label="Permalink to "Non-Technical Track: Decision Frameworks for AI Agents""></a></h1><h2 id="for-managers-pms-and-decision-makers" tabindex="-1">For Managers, PMs, and Decision-Makers <a class="header-anchor" href="#for-managers-pms-and-decision-makers" aria-label="Permalink to "For Managers, PMs, and Decision-Makers""></a></h2><hr><h2 id="framework-1-should-we-use-an-agent-for-this" tabindex="-1">Framework 1: "Should We Use an Agent for This?" <a class="header-anchor" href="#framework-1-should-we-use-an-agent-for-this" aria-label="Permalink to "Framework 1: "Should We Use an Agent for This?"""></a></h2><h3 id="the-4-question-filter-for-non-technical-stakeholders" tabindex="-1">The 4-Question Filter (for non-technical stakeholders) <a class="header-anchor" href="#the-4-question-filter-for-non-technical-stakeholders" aria-label="Permalink to "The 4-Question Filter (for non-technical stakeholders)""></a></h3><p><strong>Q1: Can a human do this task with clear step-by-step instructions?</strong></p><ul><li>Yes → Could be agentizable</li><li>No → Humans struggle too; don't agentize</li></ul><p><strong>Q2: Does the task have clear success criteria?</strong></p><ul><li>Yes → We'll know if the agent succeeded</li><li>No → You can't evaluate an agent if you can't define "done"</li></ul><p><strong>Q3: What's the cost of failure?</strong></p><ul><li>Low (wrong suggestion, formatting error) → Full autonomy possible</li><li>Medium (wrong code, wrong data) → Human review required</li><li>High (financial loss, safety, legal) → Human-in-the-loop mandatory</li></ul><p><strong>Q4: How many times does this task need to be done?</strong></p><ul><li>Once → Just do it manually</li><li>10x → Build a script</li><li>100x → Build an agent</li><li>1000x → Build a multi-agent system</li></ul><h3 id="decision-matrix" tabindex="-1">Decision Matrix <a class="header-anchor" href="#decision-matrix" aria-label="Permalink to "Decision Matrix""></a></h3><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span> Low Failure Cost High Failure Cost</span></span>
|
||
<span class="line"><span>──────────────┼─────────────────────────────────────</span></span>
|
||
<span class="line"><span>High Volume │ Build agent │ Agent + HITL</span></span>
|
||
<span class="line"><span>Low Volume │ Do it manually │ Do it manually</span></span></code></pre></div><hr><h2 id="framework-2-the-4-agent-design-patterns-andrew-ng" tabindex="-1">Framework 2: The 4 Agent Design Patterns (Andrew Ng) <a class="header-anchor" href="#framework-2-the-4-agent-design-patterns-andrew-ng" aria-label="Permalink to "Framework 2: The 4 Agent Design Patterns (Andrew Ng)""></a></h2><h3 id="_1-reflection" tabindex="-1">1. Reflection <a class="header-anchor" href="#_1-reflection" aria-label="Permalink to "1. Reflection""></a></h3><p>The agent checks its own work before presenting it.</p><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>Agent writes code → reviews own code → fixes issues → presents final</span></span></code></pre></div><p><strong>Business value</strong>: Reduces human review time by 40-60%. Catches obvious errors before they reach you.</p><h3 id="_2-tool-use" tabindex="-1">2. Tool Use <a class="header-anchor" href="#_2-tool-use" aria-label="Permalink to "2. Tool Use""></a></h3><p>The agent uses external tools to gather information or take actions.</p><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>Agent reason → call search API → read results → call database → compose answer</span></span></code></pre></div><p><strong>Business value</strong>: Agents can access real-time data, not just training data. Turns a static LLM into a live system.</p><h3 id="_3-planning" tabindex="-1">3. Planning <a class="header-anchor" href="#_3-planning" aria-label="Permalink to "3. Planning""></a></h3><p>The agent decomposes a complex request into sub-steps.</p><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>"Build a landing page" → plan steps (1. Design, 2. Code, 3. Test, 4. Deploy)</span></span>
|
||
<span class="line"><span> → execute each step → verify results</span></span></code></pre></div><p><strong>Business value</strong>: Complex tasks get done without step-by-step human instruction. The agent figures out the steps.</p><h3 id="_4-multi-agent-collaboration" tabindex="-1">4. Multi-Agent Collaboration <a class="header-anchor" href="#_4-multi-agent-collaboration" aria-label="Permalink to "4. Multi-Agent Collaboration""></a></h3><p>Multiple specialized agents work together on different parts of a task.</p><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>PM agent plans → Designer creates mockup → Frontend builds → QA tests</span></span></code></pre></div><p><strong>Business value</strong>: Each agent is an expert in its domain. Together they outperform one generalist agent at a fraction of the cost.</p><hr><h2 id="framework-3-cost-estimation-for-agent-projects" tabindex="-1">Framework 3: Cost Estimation for Agent Projects <a class="header-anchor" href="#framework-3-cost-estimation-for-agent-projects" aria-label="Permalink to "Framework 3: Cost Estimation for Agent Projects""></a></h2><h3 id="the-3x-rule" tabindex="-1">The 3x Rule <a class="header-anchor" href="#the-3x-rule" aria-label="Permalink to "The 3x Rule""></a></h3><p>A production agent costs 3x your prototype estimate. Plan accordingly.</p><table tabindex="0"><thead><tr><th>Phase</th><th>Cost Multiplier</th><th>What's Included</th></tr></thead><tbody><tr><td>Prototype (works on happy path)</td><td>1x</td><td>Single agent, one model, no safety</td></tr><tr><td>Production (handles edge cases)</td><td>2x</td><td>Retry logic, error handling, testing</td></tr><tr><td>Production+ (operational)</td><td>3x</td><td>Monitoring, security, CI/CD, budgets</td></tr></tbody></table><h3 id="quick-cost-calculator" tabindex="-1">Quick Cost Calculator <a class="header-anchor" href="#quick-cost-calculator" aria-label="Permalink to "Quick Cost Calculator""></a></h3><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>Per-task cost = (input_tokens × input_price + output_tokens × output_price) × retry_rate × overhead</span></span>
|
||
<span class="line"><span></span></span>
|
||
<span class="line"><span>Where:</span></span>
|
||
<span class="line"><span> retry_rate = 1.5 (typical for agent systems)</span></span>
|
||
<span class="line"><span> overhead = 3.0 (production multiplier)</span></span>
|
||
<span class="line"><span></span></span>
|
||
<span class="line"><span>Example:</span></span>
|
||
<span class="line"><span> Agent session: 4000 input tokens, 1000 output tokens</span></span>
|
||
<span class="line"><span> Model: Claude Sonnet ($3/$15 per M tokens)</span></span>
|
||
<span class="line"><span> </span></span>
|
||
<span class="line"><span> Base cost: (4000 × $3/M + 1000 × $15/M) = $0.012 + $0.015 = $0.027</span></span>
|
||
<span class="line"><span> With retries: $0.027 × 1.5 = $0.041</span></span>
|
||
<span class="line"><span> Production cost: $0.041 × 3.0 = $0.123 per task</span></span>
|
||
<span class="line"><span> </span></span>
|
||
<span class="line"><span> At 1000 tasks/month: $123/month</span></span>
|
||
<span class="line"><span> At 10000 tasks/month: $1230/month</span></span></code></pre></div><h3 id="cost-by-autonomy-level" tabindex="-1">Cost by Autonomy Level <a class="header-anchor" href="#cost-by-autonomy-level" aria-label="Permalink to "Cost by Autonomy Level""></a></h3><table tabindex="0"><thead><tr><th>Level</th><th>Cost/Task</th><th>Human Time/Task</th><th>Best For</th></tr></thead><tbody><tr><td>No agent</td><td>$0</td><td>30 min</td><td>One-off tasks</td></tr><tr><td>AI-assisted</td><td>$0.01-0.05</td><td>10 min</td><td>Human does the work, AI helps</td></tr><tr><td>Agent with review</td><td>$0.05-0.30</td><td>5 min</td><td>Agent works, human reviews</td></tr><tr><td>Full autonomy</td><td>$0.10-1.00</td><td>0 min</td><td>Agent works, human audits</td></tr><tr><td>Multi-agent</td><td>$0.50-5.00</td><td>2 min</td><td>Complex workflows, human supervises</td></tr></tbody></table><hr><h2 id="framework-4-autonomy-vs-risk-matrix" tabindex="-1">Framework 4: Autonomy vs Risk Matrix <a class="header-anchor" href="#framework-4-autonomy-vs-risk-matrix" aria-label="Permalink to "Framework 4: Autonomy vs Risk Matrix""></a></h2><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span> Low Risk Medium Risk High Risk</span></span>
|
||
<span class="line"><span> (suggestions, (code gen, data (financial, medical,</span></span>
|
||
<span class="line"><span> content gen) processing) infrastructure)</span></span>
|
||
<span class="line"><span>─────────────┼─────────────────────────────────────────────────────────────</span></span>
|
||
<span class="line"><span>Tier 1: │ Full autonomy │ Agent + review │ Human only</span></span>
|
||
<span class="line"><span>Simple │ │ │</span></span>
|
||
<span class="line"><span> │ │ │</span></span>
|
||
<span class="line"><span>Tier 2: │ Agent + review │ Agent + verifier + │ Human + agent</span></span>
|
||
<span class="line"><span>Complex │ │ human spot-check │ as tool</span></span>
|
||
<span class="line"><span> │ │ │</span></span>
|
||
<span class="line"><span>Tier 3: │ Multi-agent + audit │ Multi-agent + │ Not recommended</span></span>
|
||
<span class="line"><span>Strategic │ │ verifier + human │ for agents</span></span></code></pre></div><hr><h2 id="framework-5-the-build-vs-buy-decision" tabindex="-1">Framework 5: The "Build vs Buy" Decision <a class="header-anchor" href="#framework-5-the-build-vs-buy-decision" aria-label="Permalink to "Framework 5: The "Build vs Buy" Decision""></a></h2><h3 id="build-an-agent-when" tabindex="-1">Build an Agent When: <a class="header-anchor" href="#build-an-agent-when" aria-label="Permalink to "Build an Agent When:""></a></h3><ul><li>Your task is unique to your business</li><li>You need to keep data in-house</li><li>No existing tool solves the problem</li><li>You have the engineering capacity</li></ul><h3 id="buy-rent-when" tabindex="-1">Buy/Rent When: <a class="header-anchor" href="#buy-rent-when" aria-label="Permalink to "Buy/Rent When:""></a></h3><ul><li>It's a common task (customer support, code review, monitoring)</li><li>A SaaS product already exists</li><li>You don't have AI engineering expertise</li><li>Speed to market matters more than customization</li></ul><h3 id="hybrid-when" tabindex="-1">Hybrid When: <a class="header-anchor" href="#hybrid-when" aria-label="Permalink to "Hybrid When:""></a></h3><ul><li>Core task is unique, but supporting tasks are common</li><li>You can build a small agent that calls existing SaaS tools</li></ul><hr><h2 id="framework-6-measuring-success" tabindex="-1">Framework 6: Measuring Success <a class="header-anchor" href="#framework-6-measuring-success" aria-label="Permalink to "Framework 6: Measuring Success""></a></h2><h3 id="leading-indicators-week-1-4" tabindex="-1">Leading Indicators (week 1-4) <a class="header-anchor" href="#leading-indicators-week-1-4" aria-label="Permalink to "Leading Indicators (week 1-4)""></a></h3><ul><li>Task completion rate (is the agent finishing its work?)</li><li>Human review time reduction</li><li>Cost per task vs cost of human doing it</li><li>Error rate (how often does the agent need correction?)</li></ul><h3 id="lagging-indicators-month-2-3" tabindex="-1">Lagging Indicators (month 2-3+) <a class="header-anchor" href="#lagging-indicators-month-2-3" aria-label="Permalink to "Lagging Indicators (month 2-3+)""></a></h3><ul><li>Time saved per week</li><li>Quality compared to human baseline</li><li>Number of tasks automated vs manual</li><li>Maintenance burden (how often does the agent break?)</li></ul><h3 id="red-flags" tabindex="-1">Red Flags <a class="header-anchor" href="#red-flags" aria-label="Permalink to "Red Flags""></a></h3><ul><li>Agent costs more than the human it replaces</li><li>Agent requires more oversight than doing it yourself</li><li>Agent breaks every time the model updates</li><li>Team doesn't trust the agent's output</li></ul><hr><h2 id="framework-7-governance-compliance-quick-reference" tabindex="-1">Framework 7: Governance & Compliance Quick Reference <a class="header-anchor" href="#framework-7-governance-compliance-quick-reference" aria-label="Permalink to "Framework 7: Governance & Compliance Quick Reference""></a></h2><h3 id="regulatory-requirements-by-region" tabindex="-1">Regulatory Requirements by Region <a class="header-anchor" href="#regulatory-requirements-by-region" aria-label="Permalink to "Regulatory Requirements by Region""></a></h3><table tabindex="0"><thead><tr><th>Requirement</th><th>EU (AI Act)</th><th>California (SB 1047)</th><th>UK/Japan</th></tr></thead><tbody><tr><td>Human oversight</td><td>Mandatory for high-risk</td><td>Mandatory kill-switch</td><td>Voluntary</td></tr><tr><td>Audit trails</td><td>Required</td><td>Required</td><td>Recommended</td></tr><tr><td>Incident reporting</td><td>Required</td><td>Required</td><td>Voluntary</td></tr><tr><td>Disclosure (AI-generated)</td><td>Required</td><td>Required</td><td>Recommended</td></tr><tr><td>Data privacy (GDPR)</td><td>Full compliance</td><td>Similar</td><td>Different standard</td></tr></tbody></table><h3 id="minimum-viable-governance" tabindex="-1">Minimum Viable Governance <a class="header-anchor" href="#minimum-viable-governance" aria-label="Permalink to "Minimum Viable Governance""></a></h3><p>For teams not in regulated industries, start with:</p><ol><li><strong>Every agent action is logged</strong> with timestamp, agent, input, output</li><li><strong>Human review gate</strong> for destructive operations (deletes, writes, spending money)</li><li><strong>Monthly cost report</strong> per agent per task type</li><li><strong>Quarterly security review</strong> — has the agent's behavior changed?</li></ol><hr><h2 id="summary-decision-flow" tabindex="-1">Summary: Decision Flow <a class="header-anchor" href="#summary-decision-flow" aria-label="Permalink to "Summary: Decision Flow""></a></h2><div class="language- vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang"></span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>1. Can this task be automated?</span></span>
|
||
<span class="line"><span> ├── No → Do it manually</span></span>
|
||
<span class="line"><span> └── Yes →</span></span>
|
||
<span class="line"><span>2. What's the failure cost?</span></span>
|
||
<span class="line"><span> ├── High → Human-in-the-loop mandatory</span></span>
|
||
<span class="line"><span> └── Low →</span></span>
|
||
<span class="line"><span>3. What's the volume?</span></span>
|
||
<span class="line"><span> ├── <10/month → Do it manually</span></span>
|
||
<span class="line"><span> ├── 10-100/month → Build a script</span></span>
|
||
<span class="line"><span> └── >100/month → Build an agent</span></span>
|
||
<span class="line"><span>4. How many specialists needed?</span></span>
|
||
<span class="line"><span> ├── One → Single agent</span></span>
|
||
<span class="line"><span> └── Multiple → Multi-agent system</span></span>
|
||
<span class="line"><span>5. What's the budget?</span></span>
|
||
<span class="line"><span> ├── <$50/month → Use cheapest model (Gemini Flash)</span></span>
|
||
<span class="line"><span> ├── $50-500/month → Cascade routing (mix models)</span></span>
|
||
<span class="line"><span> └── >$500/month → Multi-agent with verification</span></span></code></pre></div>`,71)])])}const m=e(i,[["render",o]]);export{u as __pageData,m as default};
|