An agentic magazine about the systems that now think, write, and operate — read closely, cited properly, and built to last longer than the hype cycle.
| 01 | From the editor's desk Why a creative agency started reading footnotes. | 03 |
| 02 | The Protocol Wars MCP, A2A, ACP, ANP — four ways agents learn to talk, and the integration tax you stop paying when you pick well. | 04 |
| 03 | Agents learn to talk among themselves Inside multi-agent collaboration: actors, roles, and the structures that make a crowd of models behave like a team. | 06 |
| 04 | Required Reading Four peer-reviewable papers, each with a citation, a link, and the one line that matters. | 07 |
| 05 | Case study — Klarna's 700 What happens when a brand automates two-thirds of its support, then quietly hires the humans back. | 08 |
| 06 | The Ticker Four real-world signals from the agentic build-out, sourced and dated. | 09 |
| 07 | Notes from the floor RDLB Agentic in practice — what we shipped, what we refused, and which paper told us so. | 10 |
| 08 | Sources & colophon Every claim, linked. Plus the type, the tools, and next month. | 11 |
The agentic era arrived wearing a costume. It looked like a chatbot, then a copilot, then a dashboard full of glowing nodes. Underneath all of it is something less photogenic and far more durable: software that now reads, decides, and acts on its own — and increasingly talks to other software while it does.
We started this magazine because the conversation about that shift is loud, fast, and mostly unsourced. Every week a new post promises a "trillion-dollar" anything. Very little of it survives contact with a citation. We wanted the opposite: a monthly that reads the actual papers, reports the actual numbers, and admits when a much-celebrated deployment quietly walked itself back.
RDLB is a creative agency. So why footnotes? Because in the agentic era the brands that win will be the ones with the clearest identity and the smartest systems — and you cannot build a smart system on vibes. Taste still decides what's worth saying. But the machine that says it, at scale and on schedule, runs on protocols, collaboration structures, and governance most marketers have never had to name.
This issue maps four of them. We walk through the protocols agents use to find and talk to each other, the research on how a crowd of models behaves like a team, and a case study every operator should tape to the wall. We close with notes from our own floor — what we shipped this quarter, and the one thing we refused to ship.
The enemy this month is the easiest sale in technology right now: five chatbots in a trench coat, billed as an "autonomous workforce." Real agentic systems are humbler and more useful than that. They scale throughput. They do not, yet, scale judgment. Knowing the difference is the whole game.
— RDLB Agentic, editorial. New York. rdlbagentic.com
For two years, every agent that wanted to use a tool or call another agent needed a bespoke integration. That is the integration tax — and four competing standards are now racing to abolish it.
The problem has a name engineers use without flinching: the M×N problem. Connect M models to N tools with one-off code and you get M×N brittle integrations, each its own thing to secure, scale, and break. In November 2024, Anthropic open-sourced a fix — the Model Context Protocol (MCP) — explicitly to replace that combinatorial mess with a single way for a model to reach the systems where data lives.5
MCP is the tool layer: a JSON-RPC client–server interface for secure tool invocation and typed data exchange.1 It answers "how does one agent safely use one tool." It says nothing about how two agents, built by two companies, on two frameworks, agree to work together. That is a different protocol problem — and 2025 produced a small war of answers.
A May 2025 survey from Ehtesham, Singh, Gupta and Kumar lines up the four front-runners and, usefully, tells you the order to adopt them.1 MCP for tool access. ACP — the Agent Communication Protocol — for structured, multimodal messaging over plain HTTP, sync or async. A2A — Google's Agent-to-Agent protocol — for peer-to-peer task delegation, where each agent publishes a capability-based "Agent Card" so others can discover what it can do. And ANP — the Agent Network Protocol — for open-network discovery using W3C decentralized identifiers, so agents can find and authenticate each other with no central broker.
A second survey, "A Survey of AI Agent Protocols" from a Shanghai Jiao Tong–led group, hit its third revision in June 2025 and frames the same field along two axes: context-oriented versus inter-agent, and general-purpose versus domain-specific.2 Read together, the two papers agree on the shape of the future even where they disagree on the winners: layered protocols, decentralized discovery, and a slow drift toward what both call collective intelligence infrastructure.
A JSON-RPC client–server interface for secure tool invocation and typed data exchange. The "USB-C port" for models reaching files, APIs, and business systems. Introduced by Anthropic, Nov 2024; since adopted across the industry.5
A general-purpose protocol over RESTful HTTP — MIME-typed multipart messages, synchronous and asynchronous, runtime-independent. Session management and routing for agents that need to say more than a function call.1
Google's 2025 standard for peer-to-peer task delegation. Each agent advertises a capability-based Agent Card; others discover it and hand off work — collaboration across enterprise agents built on different frameworks.1
Open-network discovery and secure collaboration using W3C decentralized identifiers and JSON-LD graphs. A peer-to-peer model where agents authenticate and transact without a central intermediary — the groundwork for agent marketplaces.1
Protocols are not plumbing trivia — they decide your switching costs. An agent stack glued together with one-off integrations is a vendor-lock-in machine wearing a productivity halo. A stack built on open protocols keeps your data portable and your options open. The survey's advice is also a procurement checklist: start with MCP, layer ACP and A2A as you grow, and watch ANP for the day agents shop for each other.1,2
A single clever model is a soloist. The frontier of 2025 is the ensemble — many models that perceive, reason, and act together. The open question is the same one every manager faces: how do you get a crowd to behave like a team?
The most useful map of this terrain is a January 2025 survey, "Multi-Agent Collaboration Mechanisms," from Tran and colleagues at University College Cork.3 It refuses to hand-wave. Instead it characterizes collaboration along five concrete dimensions: the actors involved; the type of interaction — cooperation, competition, or the hybrid they call coopetition; the structure — peer-to-peer, centralized, or distributed; the strategy — role-based or model-based; and the coordination protocol that keeps it all coherent.
Those five knobs explain why some multi-agent systems produce insight and others produce expensive noise. A flat peer-to-peer swarm with no roles tends to loop. A role-based structure — a researcher, an editor, a critic, an operator, each with a defined job — tends to converge. The lesson is almost embarrassingly human: teams need org charts.
Why now? Because the underlying capability changed. Johannes Schneider's April 2025 survey draws the line cleanly: generative AI responds; agentic AI pursues goals, with stronger reasoning and interaction that enable autonomous behavior on complex tasks.4 Once a model can plan and call tools, putting several of them in a room stops being a parlor trick and starts being an operating model.
The same survey ends on a caution we'll echo all year: as these systems gain autonomy, the interesting risks move from "wrong answer" to "unsupervised action." Which is exactly why the boring parts — roles, protocols, governance — are the parts that pay.
Each is peer-readable, freely available, and verified live for this issue. Full citations on the Sources page.
Why it mattersThe only map that tells you the order to adopt agent protocols. Treat the phased roadmap as a procurement plan, not a reading list.
Why it mattersClassifies protocols by context-oriented vs inter-agent and compares them on security, scalability, and latency — the three questions a buyer actually asks.
Why it mattersThe five-dimension framework (actors, types, structures, strategies, protocols) is the cleanest way to design a multi-agent system that converges instead of loops.
Why it mattersThe clearest articulation of what changes when AI moves from responding to pursuing goals — and the risk register that comes with autonomy.
It is the most-cited agentic deployment on earth. It is also the most misquoted — because the story has a second half that most decks leave off.
In February 2024, Klarna and OpenAI announced an AI assistant that, within a month, had handled 2.3 million conversations — two-thirds of the fintech's customer-service chats, the equivalent work of 700 full-time agents. It matched humans on customer satisfaction, cut repeat inquiries by 25%, and resolved issues in under two minutes versus eleven. Klarna estimated a $40 million profit improvement for 2024.6,7 The internet declared the customer-service job over.
Then came the quiet part. In May 2025, CEO Sebastian Siemiatkowski told Bloomberg that cost "seems to have been a too predominant evaluation factor," and that the result was "lower quality." Klarna began a recruitment drive so customers would always have the option of a human.8 By June 2025 the framing had flipped entirely: human support would be a VIP tier — "always going to be a VIP thing," he said, like clothing stitched by hand rather than machine.9
Read uncharitably, it's a walk-back. Read correctly, it's the most honest agentic case study we have. The throughput gains were real and they stuck. What didn't hold was the assumption that throughput and quality are the same axis. Agents absorbed the volume; judgment, escalation, and the feel of being cared for did not automate as cleanly.
This is the pattern we expect to report on for years: automate the volume, keep the judgment, and price the human layer as the premium it actually is. The mistake is never "using agents." The mistake is letting the cost line write the strategy.
Map your work on two axes, not one: throughput (volume, speed, cost) and judgment (taste, escalation, trust). Hand the first axis to agents aggressively. Guard the second one on purpose. Klarna's numbers were never wrong — its single-axis scorecard was.
Dated, sourced, and free of the word "revolutionary."
MCP goes open, and the industry follows. Anthropic open-sourced the Model Context Protocol — a single standard for connecting models to tools and data — with SDKs and prebuilt servers. Within months it was being adopted well beyond its author, becoming the de facto tool-access layer for agents.5
The protocol field stops being a wild west. "A Survey of AI Agent Protocols" reached its third revision, a tell that standardization is now moving fast enough to need re-surveying every quarter. Translation: the cost of betting on the wrong integration is rising.2
Klarna re-hires the humans. The poster child of support automation publicly reinvested in human agents, recasting them as a premium tier after concluding that a cost-first scorecard produced "lower quality." The most useful correction of the year.8,9
Guardrails lag the agents. Deloitte's research finds enterprises scaling AI agents faster than the governance meant to contain them — the gap between "in production" and "under control" is now the defining operational risk of the year.10
ANP and the first credible "agents hiring agents" marketplaces; whether A2A or MCP becomes the default inter-agent layer; and the first enterprise to publish a real governance scorecard instead of a press release.
A magazine about agentic systems should run on one. Here is how this quarter's research showed up in our own build — honestly, including the part we left on the floor.
A roster with an org chart. Our system runs as twelve named agents, each with a defined role — a Keeper that guards the brand, a Listener that reads the market, an Editor that turns briefs into work, a Closer that arrives knowing the three things that matter. That is not branding garnish; it is the role-based structure Tran et al. find makes multi-agent systems converge instead of loop.3
Portability as posture. We run read-only against client systems, multi-provider by default, with the client's data staying the client's. The protocol surveys make the business case for us: open standards keep switching costs low and keep a brand from waking up locked inside one vendor's agent.1,2 We treat lock-in as the risk it is.
A content layer that acts continuously. This very magazine, and the daily insights it sits above, publish on a drip from a system of record — no manual deploy to ship a piece. That is the unglamorous definition of agentic: not a chatbot, but a system that acts on schedule, on its own, against a goal.4
What we refused. We did not ship an "autonomous marketing workforce" that replaces the judgment layer. Klarna's year taught the lesson cheaply: automate throughput, keep judgment human, and price taste as the premium it is.8,9 Our agents draft, route, and operate. A human still signs the work.
Book a 30-minute strategy blueprint call. You leave with a one-page map of your highest-cost work, sorted three ways — what to replace, what to augment, and what to leave alone — against real numbers, not vibes.
dashboardrdlbagency.com/book →