{ "docId": "019f6b8f-1773-72df-85f1-63761dd5317b", "docSlug": "jarvis-eximius-ventures-just-an-agent-away-an-ai-thesis", "documentTitle": "Eximius Ventures Just an Agent Away An AI Thesis", "authorId": "eximius-ventures-just-an-agent-away-an-ai-thesis", "authorName": null, "documentKindSlug": "deck", "documentKindLabel": "Deck", "sourceTypeSlug": "investor_relations", "sourceTypeLabel": "Investor relations", "presentationDate": null, "orientation": null, "aspectRatio": null, "pageNumber": 7, "pageCount": 22, "prevPage": 6, "nextPage": 8, "slideType": "transition", "function": "transition", "density": "dense", "nDataPoints": 0, "notes": "The slide contains a comparison of agent types (Decisioning, Agent-on-Rails, General) followed by a section transition.", "elementsJson": null, "metadataConfidence": 0.95, "imagePath": null, "slideHref": "/slides/019f6b8f-1773-72df-85f1-63761dd5317b/7", "deckHref": "/decks/019f6b8f-1773-72df-85f1-63761dd5317b", "deckJsonHref": "/decks/019f6b8f-1773-72df-85f1-63761dd5317b.json", "deckAnchorHref": "/decks/019f6b8f-1773-72df-85f1-63761dd5317b#slide-7", "components": [ { "bbox": { "h": 0.35, "w": 0.7, "x": 0.15, "y": 0.1 }, "kind": "list", "text": "2. Agent-on-Rails (\"Goldilocks\" Agent): This type balances autonomy and oversight. It has a high-level goal and some freedom to select from predefined tools and approaches, yet it remains tethered to a structured SOP or rulebook that curbs any tendency to stray off course. A typical cycle involves planning, choosing from limited actions, verifying alignment with guardrails, and then looping back to plan again. Many players (like All Hands AI and DevRev) have converged on this architecture because it preserves control while providing enough flexibility to handle varied tasks. However, building this design is more complex than creating a Router agent, as it requires weaving substantial stochasticity into a structured architecture.\n3. General Agent: At the far end are general AI agents, the so-called holy grail of agentic design. This approach, seen in early prototypes like AutoGPT, attempts to rely on the LLM's own reasoning and coding capabilities without fixed rails. Theoretically, you can implement such an agent in a simple for-loop, letting the LLM pick an action at each step. While this design sparks imagination, it remains susceptible to deviation and inconsistency. A stable, fully adaptable agent—capable of handling virtually any task with minimal guardrails—remains a long-term ambition.", "attrs": null, "subkind": "bullet", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "83b0a8ea-0d68-4593-858b-0b7fee7faf3d", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.15, "w": 0.8, "x": 0.1, "y": 0.78 }, "kind": "paragraph", "text": "OpenAI's o1 model family marks a major shift in how LLMs handle complex tasks, especially those requiring deep reasoning and multi-step logic. Traditional \"System 1\" LLMs excel at quick pattern matching but often falter on harder problems. By integrating \"System 2\" reasoning into its architecture, o1 devotes extra compute to break challenges into smaller steps, refining solutions in real-time. This approach is particularly powerful in math, coding,", "attrs": null, "subkind": "description", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "d253ba9f-8233-423f-be59-5c6cb0cd3509", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.03, "w": 0.6, "x": 0.2, "y": 0.72 }, "kind": "subtitle", "text": "What are reasoning models? What's inference time scaling?", "attrs": null, "subkind": null, "toolName": null, "toolSlug": null, "confidence": null, "componentId": "0ba5c282-c0b5-4dfa-bcd4-63575c5af56f", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.15, "w": 0.7, "x": 0.15, "y": 0.45 }, "kind": "table", "text": "Decisioning Agent: Router Agent: Constrained end of the spectrum. LLMs traverse through predetermined decision trees while most logic remains hard-coded. Example: Anterior. Key Features: Simple, limited random variation, relatively easy to code. Agent-on-Rails: \"Goldilocks\" Agent: Balances autonomy and oversight. Operates with a structured SOP or rulebook to ensure control while handling varied tasks. Example: Sierra and DevRev. Key Features: Combines planning, tool selection, and guardrails. Requires weaving stochasticity into structure; harder to build. General Agent: Autonomous Agent: Relies on the LLM's reasoning and coding capabilities without fixed rails. Example: AutoGPT. Key Features: Highly flexible but susceptible to deviation and inconsistency. Long-term ambition in agentic design.", "attrs": null, "subkind": "comparison", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "d82ba3a1-f91c-40ed-989c-6f60ba9aebfe", "frameworkName": null, "frameworkSlug": null }, { "bbox": { "h": 0.05, "w": 0.8, "x": 0.1, "y": 0.65 }, "kind": "title", "text": "Section 4: The o(h!)1 Moment - Reasoning Models and its Implications", "attrs": null, "subkind": "headline", "toolName": null, "toolSlug": null, "confidence": null, "componentId": "84f71e47-fef7-4342-b01e-11400fd2d0e5", "frameworkName": null, "frameworkSlug": null } ], "metrics": [], "tools": [], "frameworks": [], "arcBeats": [ { "to": 9, "from": 5, "beatId": "7c1dd4e3-e85c-4d9e-803d-e5c353d544ad", "arcName": "The Sparkline", "arcSlug": "sparkline", "beatName": "What Could Be", "beatSlug": "sparkline-what-could-be-2", "evidence": "The deck explores the anatomy of an agent and its implications.", "position": 1, "confidence": 0.8, "parentBeatName": "Turn", "parentBeatSlug": "turn" } ], "loops": [ { "to": 9, "from": 5, "name": "Golden Circle", "slug": "11-golden-circle", "bestFor": "Visionary leadership, brand positioning, mission statements", "matchId": "58414a40-86ef-42a4-b9c3-ad41d4b793ad", "evidence": "The deck uses a Golden Circle loop to frame the discussion around moats.", "position": 0, "objective": "Why must we look beyond tech? – Rethinking Moats", "structure": "The Why (Belief) -> The How (Process) -> The What (Result)", "confidence": 0.7, "description": "Invert the typical pitch by starting with why you exist, rather than what you do" } ], "imagePathAlt": null, "thumbSrc": null, "thumbSrcAlt": null, "locked": true }