How to read a fast-moving field
Agent tooling changes monthly; the ideas underneath change slowly. Tonight's discipline is separating the two. Standards (Part 1) and evaluation habits (Part 3) compound; framework features churn. A leader who funds the compounding layer can watch the churn with curiosity instead of anxiety.
The protocol map 15 min live
Two open standards are settling into place, and they answer different questions. MCP standardizes how a model reaches tools and context - think of it as the vertical wiring inside one agent. A2A standardizes how agents talk to each other - the horizontal wiring between agents, even agents built on different frameworks by different teams. Keep the vertical-horizontal picture and you will never confuse them again.
LiveWhy two standards, in business terms5 min▶
The split mirrors something your org already knows: connecting a system to its own databases is a different problem from getting two departments' systems to cooperate.
- MCP (model ↔ tools). One standard plug between any model and any tool or data source. Build a tool server once and every MCP-speaking agent can use it - which is why "do our tools speak MCP?" is now a foundations question, not a nice-to-have.
- A2A (agent ↔ agent). The newer layer, taught in its own Google + IBM course: a standard for agents to discover and talk to each other across frameworks - the course demonstrates interop across ADK, LangGraph, BeeAI, and Microsoft Agent Framework. Your support agent and a partner's finance agent do not need to share a vendor to cooperate.
- The leader takeaway. Standards are how you avoid marrying a framework. Tools behind MCP and agents reachable over A2A stay valuable no matter which SDK wins the year.
What is moving 12 min live
Four movements matter for your next budget cycle. None of them requires you to act this quarter; all of them reward knowing the vocabulary before your vendors use it on you.
LiveMulti-agent systems + computer use5 min▶
Multi-agent is maturing into two named shapes. The manager pattern - one agent coordinates others and synthesizes the answer - and the handoffs pattern, where agents pass a task between peers, the way a call gets transferred to the right department. Both vendors teach the same discipline you already know from session 3: maximize a single agent first, split only when tool overload or complex branching forces it. Multi-agent is a scaling decision, not a starting point.
Computer-use agents operate software the way a person does - screen, cursor, keyboard - reaching the long tail of systems that will never get an API. On the roadmap it is a watch item: powerful for legacy-heavy orgs, and exactly the kind of outward-acting capability where session 4's guardrail questions matter most.
Self-studyManaged agents + evals as procurement language4 min read▶
Managed, hosted agents. Anthropic's Managed Agents is the marker: the agent runs in a hosted, managed sandbox, so your team designs the agent and its guardrails without operating the infrastructure underneath. Expect more of the running-an-agent burden to move to vendors - which makes your golden set MORE important, not less, because it becomes the part only you can supply.
Evals are becoming procurement language. Terms from the eval literature - a golden set of tasks, pass@k (the agent succeeds in at least one of k attempts) - are starting to appear in RFPs and vendor scorecards. A buyer who can say "show me pass@k on our twelve tasks" gets a different class of answer than one who asks "is it accurate?". You learned this frame in session 5; the market is catching up to you.
The 12-month strategy 12 min live
Three lines. That is the whole strategy format: what you fund now, what you watch, what you consciously skip. Everything in it compounds; nothing in it marries you to a framework.
| Move | What | Why it compounds |
|---|---|---|
| Fund now | Tool and MCP foundations · an eval harness with a golden set · ONE pilot with its gate (your a5 recommendation). | Tools behind MCP serve every future agent. The golden set outlives every framework. The pilot converts theory into your org's own evidence. |
| Watch | A2A adoption · computer-use agents · managed/hosted agent offerings. | Real, but early. Watching costs a briefing per quarter; adopting early costs rework. Let the standards settle while your pilot earns data. |
| Skip | Framework churn-chasing - migrating because a new SDK had a good launch week. | Feature gaps between serious frameworks close within months. Switching costs are real; the eval harness makes switching possible later, when a reason exists. |
The harness beat the bet. Two teams started agent programs the same quarter. One picked the hottest framework first and built everything inside it; the other built a boring eval harness first - a golden set of real tasks with pass rates on a dashboard. A year later the framework team was locked in, negotiating from weakness. The harness team had switched stacks twice, painlessly, because every candidate had to beat the same golden set. The eval harness was the strategy; the framework was just the current tenant.
LiveClosing the leader arc4 min▶
Six sessions, one language. In a1 you learned what an agent actually is - an LLM directing its own loop - and, just as important, when not to build one. In a2 you took the anatomy apart in plain English: model, tools, instructions, and the four capabilities - tools, memory, planning, guardrails - that make an agent trustworthy. In a3 you placed workflows and agents on a single autonomy dial and learned why single-agent-first is the disciplined default. In a4 you got the risk vocabulary: seven guardrail types, prompt injection, tool risk-rating, and the triggers that bring a human in. In a5 you turned all of it on your own org - the three-criteria scan, build-buy-wait, and one pilot behind an eval gate. Tonight you placed that pilot on the roadmap. You now ask better questions about agents than most people building them - which was the point all along.
Try it yourself - this week ◐ 30 min total
- Write your org's three-line agent strategy: one line each for fund, watch, skip. If a line does not name something concrete, it is a wish, not a strategy.
- Revisit your a5 pilot recommendation with the roadmap lens: are its tools headed behind MCP? Does its golden set exist yet? Those two answers are your first two funded items.
- Book the follow-up: put the scan workshop from a5 on a real calendar with your leadership team, and bring the three-line strategy as the closing slide.
- Optional: if your builders are taking the builder track, ask them to show you the 12-task scorecard climbing - it is the same gate philosophy, running live.
Official sources covered
This track teaches the decision-maker's core of the canonical agent-building sources in business language. Certificates, videos, and graded assessments stay on the official sites (linked each session). This page covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What is the split between MCP and A2A?
Vertical vs horizontal. MCP wires one agent to its tools; A2A lets separately built agents cooperate - interop demonstrated across ADK, LangGraph, BeeAI, and Microsoft Agent Framework.
2 · Why does funding an eval harness beat betting on a framework?
The harness is the asset that compounds: it turns framework choice into a reversible decision. Teams that built framework-first negotiated from weakness; harness-first teams switched painlessly.
3 · A vendor mentions "managed agents". What are they describing?
Anthropic's Managed Agents is the marker: a hosted-sandbox product. The running burden moves to the vendor - which makes your own golden set the part only you can supply.