AI agents

What it takes to put AI agents inside a product. These notes cover the product decisions, the protocols between agents, what the tokens cost, and where a retrofit stops working.

15 articles

Start here

A dark app window with a sidebar and rows of quiet text; inside it a white tile is linked to three rows that carry white status marks, and a column of six white switches, all on, stands at the right of the window

Embedding AI agents in SaaS: 6 product decisions for 2026

Gartner expects 40% of enterprise apps to embed task-specific AI agents by end of 2026. Six product decisions separate agents that ship from agents that stall.

Three dark tiles on the left, the middle one lit white, linked to a card of skill rows pinned to a large closed dark box on the right; a white line runs from the lit tile to the lit skill row, and a row of small fading squares runs back to a lit tile beside it

A2A protocol: how to design an agent API other agents can call

A2A v1.0 shipped in March 2026. Six design steps for an A2A server: Agent Card, outcome skills, eight task states, bindings, per-call auth, versions.

Three rows of arrow shapes trailing in fading white toward one lit tile on the right

Agentic web design: what to change when agents do the buying

Gartner puts 90% of B2B buying through AI agents by 2028, over $15 trillion. Here is what changes in a product's data, feed and checkout.

A dark app window whose five lit buttons are wired through its edge to a glowing white orb on the right

WebMCP in 2026: the browser API that lets agents use your SaaS UI

WebMCP lets a web page register tools that browser AI agents call directly. Chrome 149 runs it as an origin trial. What it changes for SaaS interfaces.

A loose cloud of dim pills on the left, seven lit ones drawn by lines into ordered rows inside a window whose last rows stay empty

Context engineering: what it is and how to do it for AI agents

Context engineering decides what fills a model's context window. Get it wrong and accuracy drops 39% across multi-turn tasks, whatever the prompt says.

A stream of small white squares flowing into a block of tiles and filling it from below, stopped by a dashed white budget line, lit tiles above it

AI feature token cost: budget it before it eats your margin

Traditional SaaS runs 80% gross margin; AI-native products sit near 52%. The gap is token cost. Here is how we budget AI features so margins hold.

On a white ground, two black blocks above a waterline and a third block of equal size outlined below it, each with a range bar beside it

AI integration cost for a SaaS in 2026: real ranges

AI integration cost splits into build, run, and maintenance. Real 2026 ranges for chatbots, RAG, and agents, plus the hidden third of the bill most budgets forget.

A dim interface screen on the left wired by white lines to a lit, indented tree of nodes on the right

Designing for AI agents: building a UI the non-human user can read

AI agents read your site through the accessibility tree, the structure screen readers use. In a 2026 study, agent task success fell from 78% to 42% on a degraded one.

A lit server tile inside a countdown ring with a quarter remaining, beside eight staggered short bars fading from old to new, only the newest lit

Short-lived MCP server URLs: the TTL pattern for AI agents in 2026

1,467 MCP servers are exposed on the public internet with no auth, per Trend Micro. Short-lived URLs with a TTL of minutes, not days, close the hole.

A slide split into three panels showing a tile set into a stack, a rising line against a target, and a tile passed from one frame to another, above ten lit checks and five grey flags

AI integration agency in 2026: what their pitch should actually contain

An AI integration agency pitch in 2026 is judged on three things: what gets deployed, how it gets measured, and how it gets handed back. Ten checks to listen for, five red flags to walk out on.

On a white ground, a central tile with twelve spokes out to twelve black tiles in a ring, each spoke broken by a gap marked with a white ring

12 mistakes we see teams make building their first multi-agent ops system

Multi-agent LLM systems fail in production at 41 to 86 percent. Most of the failures trace back to twelve specific decisions teams make in the first month. Here is what they look like and how to undo them.

Above, a row of grey tiles on a thin line with one small lit tile hanging off it; below, a glowing row of lit tiles joined by a thick white line into a large lit tile

Business

What an AI-native studio actually means in 2026

An AI-native studio runs core delivery on AI agents, not on AI bolted onto hourly work. Remove the agents and shipping stops. Here is how to tell.

Five dark cards in a row, each holding a different network of lit nodes (a chain, a hub, a tree, a loop, a fan), all joined below to one white rail

Multi-agent system design patterns that survive production

Multi-agent systems cost 15 times more than chat. 40% of pilots fail in 6 months. Five design patterns decide if your AI agent team ships or stalls.

Left, a block of dark tiles with a lit tile bolted to its side and a line rising to flatten under a heavy ceiling; right, a block of lit tiles wired to a glowing centre, its line rising through a broken ceiling

AI bolted-on vs AI-native: why the retrofit hits a ceiling

Bolt-on AI hits a ceiling because the data flow is the constraint, not the model. How to tell which one you have, and when a retrofit still makes sense.

A pipe from a dark tile to a lit tile with seven leaks, each dripping a trail of dots that grow from specks to coins to a whole falling tile

7 Claude app mistakes that cost tokens, money, or the company

Seven mistakes we see in production Claude apps. Some cost tokens, some cost money, one or two can cost the company. Each one has a simple fix.

Back to Field notes