Everything You Need to Know About AI Agents | Swami Sivasubramanian | TED

Summary of technological concepts & product/solution features

AI agents: definition & behavior

• Autonomous software systems that use AI to reason, plan, and adapt toward user-defined goals.
• They can sense and interact with a digital environment, converting high-level objectives into executable steps.
• They can learn over time to improve efficiency.

How agents differ from chatbots (clear example)

• A chatbot might only propose ideas or instructions (e.g., suggest experiments).
• An agent starts with a goal and can:
  • Plan
  • Write code
  • Use tools
  • Run/construct experiments
  • Synthesize results
  • Reflect on failures
• This reduces time from weeks to minutes/hours.
• The user shifts into a trusted advisor / steering role, similar to peer reviewing.

Lowering the barrier to creating software

• Agents reduce the need for rigid specifications and large upfront project breakdowns:
  • Instead of detailing execution, you can state the goal and let agents work out the implementation.
• They can help when there are constraints like lack of skill, resources, or headcount.

---

“Milestones” required before agents can be widely trusted and adopted

1) Milestone: Make agents reach builders (software development shift)

• Agents must be practical for people building products—not just end users.
• The speaker argues agents will shift developer focus:
  • From “how to build” (implementation details like infrastructure/hosting choices)
  • To “what to build” (goal-oriented outcomes)
• AWS example (complexity reduction):
  • For EC2 hosting, there are reportedly hundreds of compute options (≈ 850).
  • Agents can hide this complexity by auto-selecting appropriate choices.

2) Milestone: Achieve trust (correctness + verification)

• Agents are acknowledged as currently imperfect, so trust is essential.
• Proposed approach: use automated reasoning to mathematically assure behavior matches specifications.
  • Automated reasoning: algorithmic proof search in mathematical logic to verify programs follow formal specs.
• Product/engineering example: Amazon Q
  • Early prototype issue: hallucinated API calls (error-prone, like an unreliable “intern”).
  • Fix: formalize API specifications into a mathematical model.
  • Add an automated reasoning “gate”:
    • When Q generates an API request, a solver verifies validity.
    • If invalid, it returns feedback to restructure the code.
  • Described as a neurosymbolic feedback loop that is transparent and enables mathematically provable correctness before execution.
  • Performance claim: verification in ~100 microseconds or less for 95% of cases.

3) Milestone: Enable anyone to build agents

• Agents should be buildable by non-programmers and across industries.
• Application example: Amazon Prime Video recap agent
  • Problem: generating a recap story can take weeks and is expensive due to manual work (story arc + scene selection).
  • Agent approach splits the work into three phases:
    1. Observation: analyze video shots/scenes to build a detailed understanding for story arc + scene selection.
    2. Reasoning: plan next steps (e.g., generate voice-over script, using observation outputs).
    3. Action: human experts review and execute final recap decisions.
  • Emphasizes human + agent collaboration to remove drudgery.

---

Guidance/framework trend (developer ecosystem)

• Agent-building frameworks are being simplified:
  • If you can write Python, you can build a useful agent.
• “Agentic cloud infrastructure” aims to move from proof of concept to production more easily.
• However, more than tools is needed:
  • Expand the builder pool beyond engineers by making interfaces familiar to business users.
  • Ensure models are grounded in real-world action (not just academic intelligence).
  • Create “worlds” (e.g., digital twin-style environments) for agents to train and improve next-generation capabilities.

---

Main speaker/source

• Swami Sivasubramanian (VP of Agentic AI at AWS) speaking.