Decision Agents
Make transparent, consistent, and reviewable decisions formalized using structured logic, sub-decisions, data dependencies for them, and predictive models with choices that are auditable, explainable, and reproducible LLMs are unreliable decision-makers, but help with analysis playing a supporting role to extract business rules from documents, summarize policy changes, or compare model versions to detect differences:
Define main decision (Approve loan?) and sub-decisions (type of asset, loan-to-value ratio, customer creditworthiness) which may rely on smaller judgments (calculate credit tier, debts, assets, or default probability).
Map all dependencies and data inputs (customer and vehicle information, etc.) = network of dependencies, where outputs of one decision feed into others.
Specify logic and, if needed, predictive ML components needed to make some decisions (probability of defaulting on a loan).
Package as a decision service deployed as an API. Other agents or workflows call this service, providing input data and receiving consistent, explainable outputs.
LLMs still play a supporting role: they can extract business rules from documents, summarize policy changes, or compare model versions to detect differences.
Source: https://www.youtube.com/watch?v=Wtpwva8t1vs&t=1031s
Agentic AI versus AI agents
Agentic AI achieves broader objectives by managing multiple agents that each focus on specific tasks.
AI Agents vs. MoE
Agents communicate at the application level and are specialized entities (data, analysis, viz. etc.) that communicate in loops: perceive, memory, reason, act, observe results, and another loop.
Experts in MoE communicate at the architecture level. MoE splits model into experts which are specialized neural nets within the same model (+ a router network), all trained together.
Source: https://www.youtube.com/watch?v=4-FH09AMsp0
Agents vs. Assistants
AI agent - performs complex, multi-step actions autonomously, makes decisions independently, learns and adapts, is proactive.
AI assistant is reactive - responds to user requests & prompts, provides info, makes recommendations, but the user makes decisions.
Bot - automates simple tasks, reactive - responds to triggers & commands, follows pre-defined rules.
Degrees of Automation
Automation without AI – human activated rules
Automation with generative AI
Low-, medium-, and high-autonomy agentic workflow
Google Antigravity
AI-powered "agent-first" IDE for software development: antigravity.google.
Google’s response to agentic coding tools like Cursor and Windsurf (fork of VS Code powered by Gemini, Claude Sonnet 4.6, GPT-OSS-120B). Antigravity treats AI agents as autonomous actors capable of planning, executing, validating, and iterating on complex engineering tasks with minimal human intervention:
Multi-agent collaboration: Multiple autonomous AI agents can work in parallel — planning, coding, testing, and browsing the web on your behalf.
Artifact generation: Agents produce verifiable deliverables (task lists, implementation plans, screenshots, browser recordings) to build user trust.
Integrated tools: Agents have direct access to the editor, terminal, and a built-in Chromium browser.
Learning: Agents can learn from previous interactions.
Antigravity 2.0 (May 19, 2026):
Desktop IDE app
Go-based Antigravity CLI tool (replacing Gemini CLI)
SDK for custom workflows
IDE vs. CLI:
IDE - multi-agent collaboration via agent manager IDE with visual feedback / artefacts and built-in browser.
CLI - single-agent terminal-native assistant, lightweight headless execution w/out background indexing and UI overhead that consume tokens silently.