Lesson 13 · Slide 1
Lesson 13 — Where Agents Are Going
A 2026 Q2 snapshot. Some bets, some open questions, one warning.
Agents
Module 6
Finale
2026 Snapshot
Lesson 13 · Slide 2
What's settled
- The agent loop is real. No serious framework rolls its own anymore — they all wrap the same shape we built by hand in L02.
- Tool use is the most useful primitive. MCP standardised it across vendors.
- Cost caps and turn caps are table-stakes. No production framework ships without them.
Lesson 13 · Slide 3
What's still unsettled
- Multi-agent orchestration shape. CrewAI's role-play vs LangGraph's DAG vs OpenAI's handoff — all three exist because none is obviously right.
- Eval methodology. LLM-as-judge is widespread but no consensus on rubrics, sampling, or how to handle judge drift.
- Memory. Every framework has some memory primitive. None has a great one for long-running agents.
Lesson 13 · Slide 4
What changed in 2025–2026
- 2025: MCP graduated from “Anthropic experiment” to industry standard. Three frameworks shipped first-class MCP support; CrewAI added it mid-year.
- 2025: “Computer Use” rebrand → “Claude Agent SDK.” General-purpose primitives, not screen-driving.
- 2026 Q1: First plausible agent-to-agent (A2A) and Agent Communication Protocol (ACP) drafts circulated. Adoption: minimal so far — multi-agent orchestration still happens inside one process.
- 2026 Q2 (now): Convergence on the same loop shape across frameworks. Specs diverge most on memory and on multi-agent coordination.
Lesson 13 · Slide 5
What to watch
- A2A / ACP adoption. If they take off in 2026 H2, the meaning of “framework choice” changes — frameworks become clients of a network of agents, not isolated boxes.
- Eval standardisation. Watch for a “promptfoo-but-for-agents” winner.
- Local-first agents. Open-weight models (Llama 4.x, Mistral Large 3, etc.) running on consumer hardware. The economics flip if you can run a Sonnet-grade agent on your laptop.
- Regulation. EU AI Act phase 2 hits in 2026 H2. Agent-specific provisions are being drafted.
Lesson 13 · Slide 6
One warning
Plan for migration from day one
The framework you pick today will not be the framework you ship in 2028. Plan for migration:
- Keep your tool implementations portable (or MCP-hosted).
- Keep your prompts in version control.
- Keep your evals framework-agnostic (they live in
_eval/, not in the agent code).
The L11 scorecard is one snapshot in a moving picture.
Lesson 13 · Slide 7
What this course did NOT cover
- Reinforcement-learning-based agents (research, not yet production).
- Fine-tuning Claude / GPT / open-weight models for agent use.
- Multi-modal agents (vision, audio, screen-driving) — beyond a brief Claude SDK note.
- Production deployment (Kubernetes, autoscaling, multi-tenant).
- Real-time / latency-critical agents (<1s response budget).
Lesson 13 · Slide 8
What this course DID cover
- Four frameworks at each one's sweet spot, hands-on.
- A reproducible cross-framework scorecard.
- The bridge from agents to MCP-hosted tools.
- Token accounting, latency budgeting, eval rigs.
- Enough to make an informed framework pick today and re-pick tomorrow.
Lesson 13 · Slide 9
Course recap
- L01 — What an agent is (and isn't).
- L02 — The agent loop, hand-rolled.
- L03 — Claude Agent SDK intro.
- L04 — Tool-heavy research agent.
- L05 — LangGraph intro.
- L06 — Graph-shaped pipeline.
- L07 — OpenAI Agents SDK intro.
- L08 — Triage + handoff agent.
- L09 — CrewAI intro.
- L10 — Multi-agent crew.
- L11 — Choosing the right framework.
- L12 — Cost, latency & evals across frameworks.
- L13 — Where agents are going (you are here).
Lesson 13 · Slide 10
Thank you.
Now go build something.
Send feedback or corrections via the repo issue tracker. Snapshot date: 2026-05-25. Re-run the L11 harness on your own task suite before making any framework decision — the numbers in this course are illustrative, your numbers are decisive.