The orchestration issue. One agent is a parlor trick. The interesting machines are crowds — a lab of agents proposing, a trusted evaluator deciding, the strongest designs surviving. We map how swarms actually work, and the coordination tax nobody prints.
Issue four. One model is a soloist. The work is an orchestra.
03
02
Orchestration
The orchestrator-worker pattern, and why a crowd with separate minds beats a single brilliant one.
04
03
Five roles, one search
Explorer, exploiter, repairer, crossbreeder, critic — a swarm is a division of labor, not a chorus of clones.
05
04
Propose, verify, evolve
The loop that lets a model be wrong safely — because a trusted evaluator, not the model, is the source of truth.
06
05
Required Reading
Four papers on swarms, evaluators, and illuminated search — each linked and verified live.
07
06
Case study — Anthropic’s 90.2%
A crowd beat a soloist by 90.2% — and spent 15× the tokens doing it.
08
07
The Ticker
Four real signals from the swarm, sourced and dated.
09
08
Notes from the floor
How RDLB Agentic runs a swarm — a trusted evaluator, a standing critic, a human gate.
10
Masthead Brave New Word is published monthly by RDLB Agentic — the operating arm of RDLB Agency. We build brand systems that think, create, and operate. We do not sell AI tools.
Editorial standard Every figure is concrete or absent. Every claim is sourced. One enemy per piece. Mechanism over metaphor. If we can’t link it, we don’t print it.
Brave New Word ✱Issue Nº 04 · August 202602
01 · From the editor’s desk
Issue four. One agent became a crowd.
For three issues we watched a single agent learn to see, to ship, and to talk. This month it stops working alone. The most capable systems of the year are not one model trying harder. They are many models, divided into roles, checked by something they cannot talk their way past.
A swarm is not a group chat. It is a structure: one agent plans and delegates, others work in parallel with their own context, and a trusted evaluator — not a vote, not a vibe — decides what is actually good. That last part is the whole game. It is what lets a model propose a thousand wrong ideas cheaply, because something honest is standing at the exit.
The payoff is real and measured. Anthropic’s research swarm beat a single strong agent by 90.2% on its own evaluation.4 DeepMind’s evolutionary systems used the same shape to make discoveries a lone model could not.1,2 The pattern works.
The enemy this month is more agents, more intelligence — the belief that stacking models automatically stacks judgment. It does not. A crowd with no evaluator is just a more expensive way to be confidently wrong, and the same Anthropic system that won by 90.2% also burned roughly fifteen times the tokens of a chat.4
So this issue is about coordination as engineering: who proposes, who decides, who keeps the cost honest. Read it, then go ask whether your "AI" is one model guessing — or a system that can be wrong safely.
A crowd with no evaluator is just a more expensive way to be confidently wrong.
— RDLB Agentic, editorial. New York. rdlbagentic.com
Brave New Word ✱Issue Nº 04 · August 202603
02 · Orchestration — Feature
The crowd beats the soloist.
The dominant shape of a serious agentic system is the orchestrator-worker pattern: one lead agent plans the work and spawns parallel sub-agents, each with its own context window and a self-contained task, whose condensed findings the lead reconciles.
Orchestrator-worker. A lead agent plans and delegates; sub-agents work in parallel, each with its own context.
Anthropic documented this directly. Their multi-agent research system used one lead "researcher" directing sub-agents, and it outperformed a single strong agent by 90.2% on their internal evaluation — the gains concentrated on breadth-first work, where several independent directions can be pursued at once.4
The minimal mental model comes from OpenAI’s Swarm, now folded into its Agents SDK: an agent is just a prompt and a set of tools, and a "handoff" is just a function that returns the next agent.5 Strip away the mystique and a swarm is a small org chart you can read — roles, tools, and the rules for passing work along.
Why split the mind at all? Because separate context windows buy you parallel reasoning and stop one long conversation from drowning in its own history. Five agents each holding one clean problem will out-think one agent holding five.
That is the upside. The next two pages are the discipline that makes it more than expensive theater: a division of labor, and an evaluator the crowd cannot fool.
Brave New Word ✱Issue Nº 04 · August 202604
Orchestration — continued
Five roles, one search.
A swarm is a division of labor, not a chorus of clones. The research lineage gives the roles their names — they are evolutionary operators with a job each.
01 · Explorer
Widen the field
Propose unlike, untried directions. Breadth first — the job is variety, not polish, so the search never collapses onto one idea too early.
02 · Exploiter
Refine what works
Take the strongest candidate and push it further — local, depth-first improvement on a lineage that is already paying off.
03 · Repairer
Fix the broken
Take a promising-but-infeasible design and make it legal against the constraints, rescuing good ideas that failed on a technicality.
04 · Crossbreeder
Combine lineages
Splice the strengths of two different winners into one candidate — crossover, the oldest trick in evolutionary search.
05 · Critic
Try to break it
Adversarial by design. The critic hunts for designs that look good to the evaluator while being fragile or gamed — the swarm’s defense against winning dishonestly.
Why a brand operator should care
A swarm without a critic optimizes for looking right, not being right. If no role is paid to attack the output, your system will quietly learn to satisfy the metric instead of the customer. Staff the critic before you scale the crowd.
Brave New Word ✱Issue Nº 04 · August 202605
03 · Innovation watch
Propose, verify, evolve.
The reason a swarm can be wildly creative without being reckless is one architectural choice: the model only proposes; a trusted, deterministic evaluator decides. The model is allowed to be wrong because something honest is the source of truth.
The loop: a model proposes candidates, a trusted evaluator scores them, the strongest survive into the next round.
DeepMind’s FunSearch named the pattern in Nature: pair a language model with an automated evaluator that guards against confabulation, then evolve the best programs across generations. It found new constructions for the cap-set problem — the largest improvement in twenty years — and better online bin-packing heuristics.1
AlphaEvolve scaled it. An ensemble — a fast model for breadth, a stronger one for depth — evolves whole algorithms under evaluator feedback. It found a way to multiply two 4×4 complex matrices in 48 scalar multiplications: the first improvement on Strassen’s 1969 result in this setting in fifty-six years.2 Same shape, larger stakes.
The lesson for anyone buying agents: ask what plays the evaluator. If the only judge of the work is the model that made it, you do not have a system — you have a confident narrator.
48
scalar multiplications for a 4×4 complex matmul2
56 yrs
since the last such improvement on Strassen2
20 yrs
cap-set gain from a model + a trusted evaluator1
Brave New Word ✱Issue Nº 04 · August 202606
04 · Required Reading
Four papers on the swarm.
From a single discovery to a scaled coding agent to illuminated search to the orchestrator-worker blueprint. Each freely available and verified live for this issue.
Romera-Paredes, B., Barekatain, M., et al. (2023) · Nature
Mathematical discoveries from program search with large language models
nature.com/articles/s41586-023-06924-6
Why it mattersFunSearch. The first time an LLM made a genuine discovery — because a trusted evaluator, not the model, decided what was true.
Google DeepMind (2025) · arXiv:2506.13131
AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery
arxiv.org/abs/2506.13131
Why it mattersThe pattern, scaled: an ensemble of models evolving algorithms under evaluator feedback — and beating a 56-year-old matrix-multiplication record.
Mouret, J.-B., & Clune, J. (2015) · arXiv:1504.04909
Illuminating search spaces by mapping elites
arxiv.org/abs/1504.04909
Why it mattersMAP-Elites. Don’t just optimize — illuminate. Keep a grid of diverse strong solutions and you see the whole trade-off surface, not one winner.
Why it mattersThe orchestrator-worker blueprint in production — with the honest economics: 90.2% better than one agent, at roughly 15× the tokens.
Brave New Word ✱Issue Nº 04 · August 202607
05 · Case study
Anthropic’s 90.2%, and the bill.
A swarm is not free intelligence. It is a trade — more capability for more cost — and Anthropic published both halves honestly, which is exactly why it is worth more than a launch-day boast.
A coordinated swarm: many small units, one formation. Power comes from the coordination, and so does the cost.
90.2%
multi-agent over a single strong agent4
~15×
more tokens than a single chat4
80%
of the performance variance is token use4
Anthropic’s research system used Claude as a lead agent directing sub-agents, each with separate context. On their internal research evaluation it beat a single agent by 90.2%, with the largest gains on breadth-first queries — the kind that benefit from chasing several leads at once.4
Then the honest part. Multi-agent runs consumed roughly fifteen times the tokens of an ordinary chat, and token usage alone explained about 80% of the performance variance. Their own conclusion: the architecture is worth it only when the task’s value clears the token bill.4
That is the buying rule in one line. A swarm is a cost decision, not a default setting. For a high-value, open-ended problem, paying 15× for a 90% better answer is obvious. For a lookup, it is waste.
The operator’s takeaway
Match the shape to the task. Reserve the swarm for breadth-first, high-value work where many directions pay off. Route the simple, single-thread jobs to one agent. The skill is not "use more agents" — it is knowing when a crowd earns its bill.
Brave New Word ✱Issue Nº 04 · August 202608
06 · The Ticker — real-world signals
Signals from the swarm.
Dated, sourced, and free of the word "magic."
2015
Illuminate, don’t just optimize. MAP-Elites proposes keeping a grid of diverse elite solutions instead of converging on a single winner — the idea that quality and variety are not enemies.3
Dec 2023
A model makes a discovery. DeepMind’s FunSearch finds new cap-set constructions — the largest improvement in twenty years — by pairing an LLM with a trusted evaluator.1
Jun 2025
Evolution writes code. AlphaEvolve evolves a procedure to multiply 4×4 complex matrices in 48 scalar multiplications — the first gain over Strassen in fifty-six years.2
Jun 2025
The crowd, measured. Anthropic reports a multi-agent system beating a single agent by 90.2% on its research eval — at roughly 15× the tokens.4
MAP-Elites: an illuminated archive. Not one winner — a grid of diverse strong designs and the trade-offs between them.
What we’re watching next month
From coordinating agents to compounding them: memory, evaluation, and the systems that get better the longer they run. That is Issue Nº 05.
Brave New Word ✱Issue Nº 04 · August 202609
07 · Notes from the floor — RDLB Agentic in practice
How we run a swarm.
A magazine about orchestration should run on an orchestrated system. Here is how this issue’s research disciplines our own thirteen-agent roster.
Orchestrator, then workers. A lead plans the work; named agents execute in parallel, each with its own context. Like Anthropic’s system, we split the mind on purpose — and only when the task is broad enough to earn it.4
A trusted evaluator decides. The model proposes; deterministic checks and read-only data decide what is true. We keep the capability separation that every result in this issue depends on — the model never gets to grade its own work.1
A standing critic. One role exists to attack the output and hunt for answers that satisfy the metric while failing the brand. It is our defense against a swarm learning to look right.
The human keeps the gate. Throughput goes to the agents; judgment stays with a person. Every action runs read-only by default, behind an approval gate, with audit-grade logs — and we price the tokens, scaling the crowd only where the value clears the bill.4
Structure from repetition: a 3D-printed lattice. A swarm’s strength is the same — simple units, disciplined arrangement.
Brave New Word ✱Issue Nº 04 · August 202610
08 · Sources — every claim, linked
If we cited it, here it is.
01
Romera-Paredes, B., Barekatain, M., Novikov, A., et al. (2023). Mathematical discoveries from program search with large language models (FunSearch). Nature. nature.com/articles/s41586-023-06924-6
02
Google DeepMind (2025). AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery. arXiv:2506.13131. arxiv.org/abs/2506.13131
03
Mouret, J.-B., & Clune, J. (2015). Illuminating search spaces by mapping elites (MAP-Elites). arXiv:1504.04909. arxiv.org/abs/1504.04909
Colophon Set in three faces echoing the RDLB system: a geometric display sans, a humanist Garamond for body and its italic payload word, and JetBrains Mono for labels. This issue leads in constellation neon — the Orchestration accent — and wears the tri-color asterisk, because a swarm is many colors holding one shape. Black, white, and the constellation palette only. Photography and diagrams generated for this issue.
Next issue · No. 05 Compounding systems: memory, evaluation, and agents that get better the longer they run. Out September 2026.
Brave New Word ✱Issue Nº 04 · August 202611
RDLB Agentic Brave New Word · Nº 04
✱
The single thing to do after reading this
Map your highest-cost 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.
Brand strategy, creative intelligence, and agentic systems — the operating arm of RDLB Agency. We build brand systems that think, create, and operate continuously.
