The inaugural issue. How software stopped answering and started acting — the reasoning loops that made it possible, the benchmarks that humble it, and the first agents to take the keyboard.
| 01 | From the editor's desk Issue one. Why a creative agency is watching the machines learn to act. | 03 |
| 02 | What is an agent, really? Perceive, reason, act, repeat — the loop that turns a model into an operator. | 04 |
| 03 | The machines learn to see Computer-use agents arrive: software that reads a screen and takes the keyboard. | 06 |
| 04 | Required Reading Four foundational papers, each with a citation, a link, and the one line that matters. | 07 |
| 05 | Case study — GAIA's 92 A benchmark where humans score 92% and a top model scored 15%. What it revealed. | 08 |
| 06 | The Ticker Four real signals from the year agents took the keyboard, sourced and dated. | 09 |
| 07 | Notes from the floor RDLB Agentic in practice — perception before action, evidence before demos. | 10 |
| 08 | Sources & colophon Every claim, linked. Plus the type, the tools, and next month. | 11 |
For two years the story was about generation — models that could write the memo. The story that actually matters is quieter: models that now read the situation, decide, and do the next thing. Generation answers. Agency acts. This magazine is about the second one.
We are a creative agency, so the obvious question is why we are opening a magazine with footnotes instead of a moodboard. The answer is that in the agentic era the brands that win will pair the clearest identity with the smartest system — and a system is not a vibe. It is a loop: perceive, reason, act, check, repeat. Taste decides what is worth doing. The loop decides whether it gets done at scale and on schedule.
This inaugural issue maps the foundations. What an agent actually is, underneath the marketing. How a model learned to interleave thinking and doing. And what happened the first time we measured these systems on real-world tasks instead of demos — a humbling we think every operator should sit with.
The enemy this month is the most common artifact in the industry: the demo that never ships. A scripted clip of an agent booking a holiday is not a system; it is theatre. The work is in the unglamorous gap between a thing that works once on stage and a thing that works on Tuesday, unattended, on your data.
We will spend the year in that gap, reading the papers and reporting the numbers. Welcome to Brave New Word.
— RDLB Agentic, editorial. New York. rdlbagentic.com
Strip away the marketing and an agent is a loop, not a personality. It perceives a state, reasons about it, takes an action through a tool, observes what changed, and goes again — until the goal is met or the budget runs out.
The clearest map of this terrain is a September 2023 survey, "The Rise and Potential of Large Language Model Based Agents," from Xi and twenty-odd colleagues — later carried in Science China Information Sciences.1 It frames an LLM agent as a brain (the model) wired to perception and action, and catalogues how such agents are built, deployed, and combined into societies. The useful part is the insistence that an agent is an architecture, not a prompt.
The mechanism that makes the loop work has a name and a paper. "ReAct," from Yao and colleagues in 2022, showed that interleaving reasoning traces with actions beats doing either alone: the reasoning lets the model plan and handle exceptions, while the actions let it reach out to a tool or environment for facts it does not have.2 Think aloud, then do something, then look at the result, then think again. Almost every agent shipped since is a variation on that rhythm.
One more ingredient turns a loop into something that improves: memory of its own mistakes. "Reflexion" (Shinn et al., 2023) reinforces an agent not by retraining weights but by having it write a short verbal critique of what went wrong and keep it in an episodic buffer for the next attempt.3 Cheap, legible, and effective — self-correction as a sticky note rather than a gradient step.
Put the three together — architecture, the reason-act loop, and reflective memory — and you have the minimum viable agent. Everything fashionable in 2025 is an elaboration of it.
Ingest the situation — a document, a screen, an API response, a database row. The richer and more grounded the perception, the less the model has to hallucinate.1
A reasoning trace: what is the goal, what is missing, what action gets me closer. ReAct's insight is that writing this down, then acting, beats leaping straight to an answer.2
Search, run code, query a system, click a button. The action is where the agent touches the world — and where it can do real work or real damage.
Observe what changed, critique the attempt, and carry the lesson forward. Reflexion stores that critique as plain text for the next try — improvement without retraining.3
The loop is also a control surface. Each move is a place to insert a check: a perception you can audit, a reasoning trace you can read, an action you can gate, a reflection you can log. Buy or build agents that expose these four moves and you keep oversight. Buy a black box that only shows you the final answer and you have bought theatre with a login.2,3
For a decade, software you didn't integrate with was software an agent couldn't use. Then, in the space of three months, the major labs taught models to do what people do when there's no API: look at the screen and take the keyboard.
In October 2024, Anthropic shipped computer use — a model that moves a cursor, types, and clicks by reading screenshots, no bespoke integration required.5 In December, Google DeepMind previewed Project Mariner, an agent that browses and acts inside the browser on your behalf.6 By January 2025, OpenAI's Operator put a computer-using agent in front of consumers, booking and ordering through a real web browser.7
This is a genuine unlock. Perception through pixels means the long tail of software without APIs — most of it — becomes addressable. The same shift shows up in the research: GAIA, the benchmark we anchor this issue on, deliberately mixes reasoning, multimodal understanding, and tool use, because real assistance requires all three at once.4
It is also where the honesty begins. Seeing a screen is not the same as reliably operating it, and the early computer-use agents are slow, brittle on unfamiliar layouts, and prone to confident wrong clicks. The capability is real; the dependability is a work in progress. Which is exactly the gap the next page measures.
The foundations. Each is freely available and verified live for this issue. Full citations on the Sources page.
Why it mattersThe field's base map: an agent is a brain wired to perception and action, not a clever prompt. Start here before you buy anything.
Why it mattersThe loop, named. Interleaving a reasoning trace with tool actions is still the backbone of nearly every agent shipped since.
Why it mattersSelf-correction without retraining — the agent writes its own post-mortem and reads it next time. Cheap, legible improvement.
Why it mattersThe reality check: real-world assistant tasks where humans score 92% and a top model with plugins scored 15%. Demos lie; this doesn't.
The most useful agentic result of the era is not a triumph. It is a gap — and it is the number every operator should tape to the wall before approving a deployment.
GAIA, introduced in late 2023 by a team including Meta's Yann LeCun and Hugging Face's Thomas Wolf, asks questions that are conceptually simple for a person but require an assistant to chain reasoning, browse the web, read a PDF or image, and use tools to answer.4 Booking-clerk stuff. Humans got 92%. GPT-4 with plugins got about 15%.
That gap is the whole story of this issue. The models could reason beautifully in isolation; what they could not yet do was string perception, tool use, and grounding together reliably enough to finish a real errand. Intelligence was not the bottleneck. Execution was.
This reframes how to read every flashy demo. A model that aces a closed benchmark can still flounder the moment the task touches the messy, multi-step, real world. The diagnostic is not "is it smart?" but "can it finish the errand, unattended, more than once?"
The good news in the bad number: GAIA is exactly the kind of yardstick that turns hype into a roadmap. You cannot improve what you will not measure honestly — and a benchmark that humbles your system is worth more than a demo that flatters it.
Before you approve an agent, ask for its score on tasks that look like your work, not a vendor reel. If the only evidence is a demo, you are buying the 15% and being shown the 92%.
Dated, sourced, and free of the word "magical."
Anthropic ships computer use. A model that operates a desktop by reading screenshots and moving the cursor — no per-app integration. The long tail of API-less software becomes, in principle, addressable by an agent.5
Google previews Project Mariner. A browser agent from DeepMind that navigates and acts inside the page on your behalf — research prototype first, wider release later.6
OpenAI puts Operator in front of users. A computer-using agent that books and orders through a real web browser — the consumer debut of "the agent takes the keyboard."7
The benchmarks get honest. GAIA reframes evaluation around real, multi-step errands — and posts the gap (92% human vs ~15% model) that the demos had been hiding.4
From "can it see?" to "can it ship?" — the move from perception demos to production work, and the first honest reliability scores. That is Issue Nº 02.
A magazine about agents should be built by one. Here is how this issue's research shows up in our own system — including the discipline it imposes on us.
We build the loop, not the puppet. Our agents expose the four moves — perceive, reason, act, reflect — so a human can audit each one. That is the ReAct rhythm by design: a reasoning trace you can read, an action you can gate.2
Perception before action. Most agent failures are perception failures wearing an action costume. We invest first in grounding — clean, current, well-structured context — because GAIA's gap is mostly a grounding-and-tool-use gap, not an IQ gap.4
Reflection is logged, not hidden. When an agent gets something wrong, the critique is written down and kept — Reflexion as operating practice — so the system gets better in the open rather than failing the same way twice in the dark.3
What we refuse. We do not ship the demo that never ships. Nothing goes live on a client's brand until it clears tasks that look like their real work. The reel is not the system.
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→