About This Talk
A conversation about agentic architectures and the new era of AI-driven decision-making.
Key Takeaways
- Your agentic architecture matters more than the number of agents
- Lower intelligence LLMs can outperform groups of higher intelligence models with the right composition
- These tools allow you to stress test your agentic roadmap to reduce risk
1
Sequential Interdependence
Is the task fundamentally interactive, requiring a sequence of actions where the result of one step dictates the input for the next?
Sequential Interdependence
Is the task fundamentally interactive, requiring a sequence of actions where the result of one step dictates the input for the next?
Think of it like cooking a steak. You can't just set a timer and walk away—you need to touch it, check the sear, flip it based on what you see, and pull it off when it feels right. Each decision depends on what happened in the previous step. That's sequential interdependence: the task can't be solved with one instruction upfront.
Agentic tasks require sustained multi-step interactions with an external environment. If a high reward cannot be achieved by a one-shot policy, sequential interdependence is required. This differentiates agentic tasks from static benchmarks solvable through single-shot reasoning.
2
Partial Observability
Does the system need to actively gather hidden information or use external tools before it can reach a solution?
Partial Observability
Does the system need to actively gather hidden information or use external tools before it can reach a solution?
Imagine you're a detective at a crime scene. You don't see the whole picture right away—you have to interview witnesses, dust for prints, and follow leads. Each clue you uncover changes where you look next. That's partial observability: you can't solve the case from your desk; you have to go out and actively discover the hidden pieces.
The task must involve iterative information gathering under partial observability. Critical state information is hidden and must be acquired through active querying or tool use.
3
Adaptive Strategy Refinement
Must the system change its strategy, revise its beliefs, or adapt its plan based on feedback from the environment or tool failures?
Adaptive Strategy Refinement
Must the system change its strategy, revise its beliefs, or adapt its plan based on feedback from the environment or tool failures?
Picture a road trip where GPS reroutes you. You planned to take the highway, but there's a wreck ahead. Your GPS doesn't stubbornly insist on the original route—it adapts, finds a new path, and gets you there anyway. That's adaptive strategy refinement: when the world pushes back, you pivot instead of failing.
The agent must perform adaptive strategy refinement. It must update its internal beliefs and revise its strategy based on new evidence obtained through interaction and environmental feedback.