Column
AI 2026: First-Half Review, Second-Half Forecast
Five essays — from what actually happened, to what comes next
Information is cheap. What is expensive is the ability to process it.
This column sets out to do something plain: take a handful of things that actually happened in AI Agent land during the first half of 2026, explain them properly, and then reason forward into what the second half most likely holds.
I don’t chase hype. Hype depreciates the day it lands, and chasing it only leaves you busier and emptier. What I chase are signals — the things that, once you see them clearly, change what you do next. These five essays are the signals left standing after repeated filtering against a large number of frontier system releases, primary sources, papers, and benchmarks:
One — the ceiling on information automation. Point an AI at a field and tell it to track everything: how far does it actually get, and where does it jam?
Two — the shift in interaction. Agents are moving from “you prompt it” to “it prompts you.” What that step actually costs.
Three — the collapse in cost. Once open models get roughly an order of magnitude cheaper, how large a fleet can one person really keep running?
Four — the engineering of trust. What it takes to hand work to an agent nobody is watching: evals, guardrails, and a human in the loop.
Five — the map of opportunity. The closer: where the red ocean has filled in, and what blue ocean is left.
The five interlock, but each stands on its own. Read 1 → 5 and you’ll get the full path: from observation, to forecast, to what you should actually bet on.
Contents
Best read in order-
Where an AI That Tracks a Whole Field For You Finally Gets Stuck
If you hand an AI the job of tracking a whole field for you, how far does it actually get? This essay takes apart the AI information pipelines that exploded in the first half of 2026: where each of the three technical routes stalls, how much three-stage dedup can cut and what it can never cut, and why compounding error makes long chains drift by construction. Access has been cut to zero, but the output usually stops at a more polished bookmark folder. The real ceiling is structural. AI can know things for you; it cannot judge for you. Includes a build checklist you can copy directly, a list of anti-patterns, and three forecasts for the second half.
-
When the AI Agent Starts Prompting You, What Has Actually Changed
The counterintuitive signal of the past six months is that a growing number of AI agents no longer wait for your instruction. They read your context and prompt you instead. This essay places the proactive agent inside the third migration of interaction paradigms, from CLI-style to chat-style to proactive, takes apart the three foundations it needs, persistent memory, an agent runtime, and event triggers, offers an interruption budget model and a speak-up threshold you can apply directly, and makes a forecast for the second half, proactive will become a buzzword, but a wave of these products will die first, because the cost of interruption is badly underestimated and the tradeoff between trust and interruption is the real dividing line.
-
Open Models Got Roughly 90% Cheaper. How Big an Agent Fleet Can One Person Afford?
An agent is a workload that burns tokens, and cost is its law of physics. By my own estimate, at comparable intelligence open models are now roughly 90% cheaper, taking a single agent task from a few dollars down to a few cents. For the first time, one person can genuinely afford a fleet of agents running around the clock. This essay works through the arithmetic and one ASCII diagram to explain the token cost collapse, why multi-model routing went from an optimization to the default architecture, and the hidden costs that cheap tokens quietly leave on your plate.
-
How Do You Get to the Point Where You Trust an AI Agent Nobody Is Watching?
The real bottleneck in handing work to an agent was never whether it can do the work, it is whether you dare use what it hands back. This essay takes apart the trust stack for unattended AI agents: guardrails that authorize before the action fires, regression evals that catch drift, and a morning HITL review. It uses the hard math of compounding error to explain why 95% per-step accuracy leaves you around a third after twenty steps, sorts guardrails into four tiers by reversibility, shows how to cold-start an eval set from your own incident log, and ends with a morning checklist and a list of anti-patterns. Trust does not come from the model behaving. It comes from engineering.
-
The Red Ocean Is Full: Where the Blue Ocean for AI Agents Is in Late 2026
In late 2026, general-purpose and horizontal AI agents (coding, support, sales, scheduling) are a red ocean that foundation model labs can replicate within a single product cycle. The real blue ocean sits in two places: vertical, regulated workflows where someone takes end-to-end accountability, and infrastructure built for agents themselves (runtimes, payment rails, evals). This piece explains why feature moats cannot hold, what a real moat looks like (proprietary data flywheels, domain depth, end-to-end accountability), and offers a red-blue test table, a direction self-check, and an anti-pattern list, with separate bets for solo builders and founders. It closes out the entire 2026 review-and-forecast column.