The agentic era, explained one idea at a time.
the achilles angle · the columnAI, made legible, with a little wit.
Achilles and the cast explain the agentic era one idea at a time. Short reads, real insight, zero jargon.
The orchestration issue. One agent is a parlor trick; the interesting machines are crowds. The orchestrator-worker pattern, propose-verify-evolve, and Anthropic's multi-agent system that won by 90.2% — at 15× the tokens.
One useful note on brand systems, agentic workforces, and staying unmistakably yourself. No spam, no filler, unsubscribe anytime.
Pilots run under conditions a rollout never gets. The gap is not the model. It is memory, enforceable rules, approval, and logs.
Frontier weights are a download now. The thing still worth fighting over is where demand arrives.
When throughput triples, review becomes the constraint. Tiering, enforceable rules, and logs are what keep the gate from collapsing.
Everyone is selling agentic systems now. Here is ours, running in public, with receipts.
Every piece of brand work starts as a brief. Treat the brief as the unit—structured, versioned, enforceable—and throughput compounds without the drift.
Founder judgment is the scarcest asset in the company. The ceiling forms when it is spent reviewing execution. How an approval gate changes the math.
Most agent risk is access design, not model behavior. Read-only connectors, one approval gate, and audit-grade logs make autonomy safe to run.
This column is produced by an agentic system. Here's the machinery that keeps it from wandering.
This week's agents can run for hours unsupervised. The feature that actually matters is how easily you can stop them.
Open or closed models is a routing decision, not an identity. Model-agnostic architecture matters more to a brand system than the model you pick this quarter.
A lone agent is a tool. The value is in the coordination around it.
Most brands judge an agentic system by its first asset. The real meter is the ninety-day slope: throughput climbs three to five times while cost stays flat.
Thirteen agents. 44,000+ production runs in 63 days. Under $50 in model spend. What collapses when execution gets cheap, and what stays expensive.
Every approved brief and killed concept is a decision. Most brands lose them. A system that remembers turns each one into compounding leverage.
Where marketing automation stops, what persistent memory changes, and three honest checks that tell you which system your team actually needs.
A working definition of agentic marketing, the four-part mechanism behind it, and the tool stack plus manual assembly it retires.
The models are a commodity now. What you point them at isn't.
Audit-grade logs turn agent output into evidence: what ran, what changed, who approved it. Trust in an agentic system is a record, not a feeling.
Most agentic projects do not fail on technology. They fail on adoption. A capability nobody trusts or uses returns nothing. The last mile is human.
Cheaper content is the smallest reward an agentic system offers. The real prize is operating leverage: more output per unit of judgment, without more headcount.
Every handoff between people loses context and adds wait time. How an agentic system carries context end to end and removes the coordination tax.
Vendor lock-in feels free at signing. The bill arrives when the model changes, the price moves, or you want to leave. Portability is the hedge you buy now.
Buyers now ask an assistant before they visit your site. If a model cannot read your brand clearly, it recommends someone it can. That is a systems problem.
The human approval gate looks like friction. Priced correctly, it is the cheapest control in an agentic system and the one that protects every other line.
When an agent fails, the fix is rarely a smarter model. It's usually a fuller brief.
When people leave, your brand's accumulated judgment leaves with them. An agentic system holds the decisions, briefs, and voice rules in place.
A demo shows it worked once. A test set shows it works when you're not watching.
One capable agent still guesses. A swarm proposes, checks, and keeps the best — the orchestrator-worker pattern, and how to run it without the bill running you.
A brand voice document describes the voice. It does not enforce it. Here is how voice becomes rules a system applies to every output, every time.
Most agentic systems get scored on the year-one fee. The questions that predict whether one will serve your brand are about exit, control, and memory.
Not every task belongs on an agent. A practical order of operations for what to automate first in a brand system, and what to keep firmly human.
Capital, talent, and culture are temporary edges. Real advantage is structural: proprietary data and a system that compounds it. A Columbia view.
A look at what an agentic brand system completes overnight, why a system beats a tool, and where the founder still owns the final yes.
Volume is cheap; taste is scarce. How a written standard and a human approval gate let a brand scale output three to five times without lowering the bar.
Agentic pilots fail on operating problems, not model quality. How read-only access, audit-grade logs, and a human gate turn a demo into a rollout.
Brand work moves at the speed of its briefs. Treat the brief as the asset and an agentic system can run it at scale without losing your voice.
The most expensive work in a founder-led brand is judgment, and it lives in one head. How an agentic system encodes it and lifts the throughput limit.
The dashboard was a workaround for the fact that humans couldn't talk to systems. As that changes, where does brand work actually happen?
When a model can read your data directly, a dashboard that only summarizes it becomes a tax. What replaces the analytics layer for brand operations.
General models know everything and nothing about your brand. The defensible advantage is the data only you have — captured, structured, and recalled.
If one vendor's policy change can break your brand operation overnight, you have a dependency, not a system. How to design the risk out.
The model layer is a decision, not a default. How to weigh frontier performance against control — and why the choice should live with you, not a vendor.
Agents beside your systems, never inside them. Why sovereign runtime and scoped connectors are what let legal sign off in week one.
The flagship engagement, week by week: signal in, judgment through, work out — wired into your business and owned by your team at the end.
Most brands pay twice for the same work — once for the team, again for tools nobody operates. How to find the number, and what a system changes.
Models change quarterly; your brand does not. The architecture that lets the system upgrade automatically while the brand layer holds.
Autonomy is not the goal — governed throughput is. How approval gates, audit logs, and scoped access keep an agentic system worth trusting.
Two to three weeks, no retainer: mapping your highest-cost work, scoring it for agentification, and handing you the build plan — whether or not you build with us.
AI agents now research and shortlist brands before a human loads a page. What they read, what they skip, and how to be the safe recommendation.
Voice drift is not a talent failure — it is what happens when consistency depends on memory. The mechanism for sounding like yourself at scale.
Agents are roles, not chatbots: named operators with schedules, handoffs, and outcomes. How a 12-operator workforce divides real brand work.
Not a tool, not a retainer — a governed system of agents that holds your voice, reads your market, and produces continuously. The mechanics, plainly.
