Agents (LLM Agents)

Definition

Systems that place an LLM at their core as the controller to perform multi-step planning + tool use + environment interaction. The ability to carry out sequential tasks beyond a single response.

Representative patterns:

• ReAct (Reason → Act → Observe loop)
• Reflexion (self-critique + retry)
• Tool calling (function calling, MCP)
• Code generation as control flow

Why It Matters

• Squarely at the center of personal interests (1.5x weighting)
• The thickest current in the 2026 AI industry — nearly every frontier lab has an agent line
• The inflection point in the LLM's transition from "text generator" to "general-purpose worker"

State of the Art (2026-05)

Industry side

• Claude Opus 4.7 — Anthropic, emphasizing agents/multi-step
• GPT-Rosalind + Codex — OpenAI, OpenAI's software engineering agent
• Muse Spark (1.0 / 1.1) — Meta MSL, emphasizing "personal"
• Mistral Medium 3.5 + Vibe remote agents — European camp
• Google DeepMind AlphaEvolve — coding agent

Agent Platforms & Infra (2026-06)

• Microsoft Build 2026 Agent Stack (confirmed in the 2026-06-02 keynote): the entire agent execution stack launched simultaneously.
  • Windows Agent Framework 1.0 — MIT-licensed open source; supports Windows 11/365/Azure Arc; built-in human approval queue (actions requiring permissions must be approved by a human); model-agnostic
  • Azure Agent Mesh — a federated multi-agent execution platform spanning on-prem Windows/Windows 365/Azure Arc edge; uses the same API as local; automatic routing based on latency/GPU availability; GA target Q4 2026
  • GitHub Copilot App — an agent-native standalone desktop app (macOS/Windows)
  • VS Code multi-agent GA — on the day of Build
  • Copilot Workspace GA (Enterprise) — an autonomous agent for bug fixes/writing tests/opening PRs
  • Azure AI Foundry — first-class support for Claude, Mistral, Llama 4, DeepSeek → multi-model agent orchestration
  • Project Polaris — MoE coding agent, GitHub Copilot default model GA 2026-08
  • MAI-Code-1 / MAI-Code-1-Flash — 5B class, GA across all Copilot tiers on the day
  • MAI-Thinking-1 — 35B active, specialized for reasoning/orchestration → Microsoft (source)

Non-Developer Agents: Codex for Every Role (2026-06-02)

OpenAI announced role-specific plugin expansion of Codex (source):

• 6 role-specific plugins with 62 popular apps + 110 skills (Analyst, Marketer, Operator, Designer, Researcher, Investor/Banker)
• Codex Sites (preview): creates interactive hosted web apps from text → dashboards, planners, review workspaces
• Non-developers: 20% of Codex users, growing 3× faster than developers
• Coming: Corporate Finance, PE, Marketing Strategy, Strategy Consulting, Legal plugins
• Significance: "AI agent" is crossing the developer ↔ non-developer boundary; software engineering agents generalizing to knowledge work. Most capable agent (Codex) now positioned as universal knowledge worker tool.

→ OpenAI

Multi-Agent Harness Design for Long-Running Tasks (Anthropic, April 2026 —

Anthropic Engineering Blog describes a three-agent architecture enabling multi-hour autonomous coding/frontend design sessions (source):

• Planner — expands a short product prompt into full spec
• Generator — builds the application
• Evaluator — uses Playwright MCP to test behavior against contracts (separating creation from critique)
• Context reset technique — completely clears context window at overflow, passes structured handoff (state + next steps) to fresh agent
• Results: solo run $9/20 min vs. full harness $200/6 hr → significantly more complete product
• Related: Anthropic 2026 Agentic Coding Trends Report (Jan 2026) — "delegation gap": developers delegate only 0-20% of tasks; harness design addresses quality problem of longer autonomous runs

Significance: Context reset + structured handoff is an emerging pattern for scaling agent runs beyond single context windows — enables production-quality outputs without manual re-intervention.

→ Anthropic

Computer Use (GUI Agents)

• Anthropic + Vercept (Feb 2026): the Vercept acquisition strengthens Claude computer use. OSWorld <15% (2024-Q4) → 72.5% (2026-05). Cloud-hosted MacBook remote-control architecture. → Anthropic (source)
• Browser and desktop GUI manipulation is emerging as an important execution environment for agents — general-purpose computer use beyond code execution

Research side

• Agentic Reinforcement Learning — the RL learning direction
• Embodied Agents — extension into the physical world
• MCP (Model Context Protocol) — Anthropic standard → broad adoption (default Microsoft 365 Copilot integration 2026-01)

Sub-concepts

• Agentic Reinforcement Learning — the RL aspect
• Embodied Agents — the physical-world aspect
• tool-use, mcp, reflection separate pages TBD

Persistent Memory: OpenAI Dreaming V3 (2026-06-04)

OpenAI released Dreaming V3 — a background memory synthesis architecture that represents the most significant advance in agent/user memory since the original ChatGPT memory rollout (source):

• Mechanism: Background process continuously synthesizes salient facts, preferences, and temporal context from all conversations — no explicit user instruction required
• Temporal awareness: Auto-updates time-sensitive memories (e.g., past events are revised from future tense to past tense)
• Replaces: The explicit "saved memories" list as the standalone memory foundation
• Performance: Factual recall 41.5% (2024) → 82.8% (2026); preference/time-sensitive accuracy in low-70s
• Compute: 5× reduction in memory-related inference compute → enables Free-tier rollout
• Rollout: June 4, Plus/Pro US first; Free and global to follow

Disambiguation: OpenAI "Dreaming" (personalization memory for ChatGPT users) vs Anthropic "Dreaming" (agents reviewing past sessions to form procedural memory for self-improvement — see Claude Managed Agents). Same name, different mechanism.

Agent relevance: Persistent user memory is a foundational requirement for continuity in long-running agents. Dreaming V3 solves this at the product layer for consumer ChatGPT; the architectural pattern (background synthesis vs. explicit storage) is the key technique.

→ OpenAI

Google ADK 2.0 GA + Agents CLI (2026-06-30)

Google's ADK reached GA with graph workflows (fan-out/fan-in, loops, state management, human-in-the-loop) and a collaborative Task API for agent-to-agent delegation. New Agents CLI covers the full lifecycle in one tool (scaffold → evals → deploy → observability → publishing). Compatible with Claude Code, Cursor, Gemini CLI. Addresses the key eval gap (89% teams have observability, only 52% have evals). → Google ADK (Agent Development Kit), Google DeepMind (source)

Open Problems

• Long-horizon stability — drift / failure accumulation in 50+ step tasks
• Verification — can an agent verify the results of its own work?
• Memory — context limits vs. cumulative learning (Dreaming V3 offers a consumer-layer solution)
• Cost — token cost per agent run vs. gain
• Safety — side effects of autonomous action

Key Papers / Events

• Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent — 2026-06-30: horizon scaling beats parameter scaling for agentic tasks (35B MoE matches 1T-class models via 45K-token trajectories)
• Self-Distilled Agentic Reinforcement Learning — 2026-05 latest

Related Concepts

• Agentic Reinforcement Learning
• Embodied Agents
• Reasoning Models

Notable Statements

• Andrej Karpathy (2026): "rapidly shifting from 80% manual + 20% agents to 80% agent coding + 20% edits"
• Jim Fan: "LLM acts as 'prefrontal cortex' that orchestrates lower-level control APIs"
• Andrej Karpathy: "it's hard to imagine what creating software at the end of 2026 will look like" (via Sam Altman echo)

Open Debates

• Agent = LLM + scaffolding vs Agents require separate training — part of academia (the agentic RL camp) argues the latter. Part of industry argues the former is sufficient.
• No-gradient orchestration (Jim Fan's position) vs end-to-end neural control (Sergey Levine's camp) — a fork in the learning paradigm.