Guides
Field notes on agent interfaces
Working guides to agent interfaces: the protocols that wire agents into software, and the UX patterns that keep humans in command.
OpenAI's app directory quietly became a plugin standard
In ten months, OpenAI's ChatGPT App Directory went from a proprietary store to a directory built on a cross-vendor plugin spec it helped write.
2026-08-14
Multi-agent supervision: the fleet view arrives
Claude Code, Devin Desktop, and Cursor all shipped a dashboard for watching several agents at once this year — and independently landed on the same shape.
2026-08-13
MCP Apps: UI enters the protocol, officially
Tool UI went from proposal to formal MCP extension in eight months. The mechanics, the client list, and what security researchers are already flagging.
2026-08-12
Agent Plugins: a shared format for skills and MCP
Six companies agreed on one plugin format for AI agent skills and MCP servers. Anthropic, whose specs it packages, isn't among them.
2026-08-11
OS-level agent workspaces: the OS as the permission gate
Windows 11 and Apple are both moving agent permissions into the OS and the app-API layer. Here's how the two mechanisms actually differ.
2026-08-11
ACP one year in: the LSP for coding agents
Zed's Agent Client Protocol went from a Gemini CLI side project to JetBrains' default and a public agent registry — with v2 now in draft.
2026-08-08
Cost transparency: what to show before an agent spends
Token spend and real-money spend are tracked by completely different systems today, and neither shows you the number before you say go.
2026-08-08
The empty prompt problem: what to show first
A blank text box asks new users to recall what an agent can do before they have seen it work. Here is what gets people to a first real result instead.
2026-08-06
llms.txt in 2026: adoption is up, readership still isn't
More sites ship llms.txt than ever, but AI crawlers still ignore it — while Chrome's own Lighthouse tool just started auditing for one.
2026-08-06
Trust calibration: why agents should show their doubt
Agents sound equally confident whether they're right or guessing. Calibrated uncertainty is a design problem, not just a training one.
2026-08-05
Streaming progress: what to show while an agent works
Not every tool call earns a line on screen. What shipped agent products actually stream, what they hold back, and where the line falls.
2026-08-04
Agentic browsers in 2026: what actually shipped
Atlas dies August 9, Comet went free and cross-platform, and Edge and Brave folded agentic browsing into browsers people already use.
2026-08-04
Computer use in 2026: three different bets on the screen
Anthropic, OpenAI, and Google each shipped computer-use agents, and disagree on whether the right scope is the desktop, the app, or the browser.
2026-08-03
Handoff design: escalation without losing context
When an agent hands off to a human, the failure mode isn't the escalation — it's the context that doesn't survive it.
2026-07-31
AP2 and x402: two answers to how agents pay
Google's mandate-based AP2 and Coinbase's HTTP-native x402 solve different halves of agent commerce, and increasingly plug into each other.
2026-07-31
MCP goes stateless: what the 2026-07-28 spec changes
MCP's biggest spec revision yet ships today: sessions are gone, replaced by explicit state. Nothing breaks immediately — but plan the migration.
2026-07-30
When a long-running agent should ping you
A three-tier framework for agent notifications, presence-aware suppression, and what separates a useful ping from noise.
2026-07-30
AG-UI in 2026: fast adoption, no foundation yet
CopilotKit's frontend protocol landed in AWS Bedrock and raised $27M this year, but it's still the one major agent protocol without a foundation home.
2026-07-27
Undo for agents: reversibility as the safety net
Approval gates ask permission before an action. Undo is the backstop for what happens after — and coding agents are quietly building it out.
2026-07-27
What to log so an agent's audit trail survives an incident
Agent audit trails need more than request logs — they need reasoning, authorization, and identity. What regulators and cloud vendors now require.
2026-07-25
Stopping an Agent Mid-Task Without Corrupting State
Every serious agent surface needs a stop button. Most don't have one that reliably works, and the ones that do show why it's hard.
2026-07-24
Who Approves What: Enterprise Agent Governance in 2026
Agent identity platforms from Microsoft and Okta went GA this year. Adoption is still outrunning the controls meant to govern it.
2026-07-24
The agent graveyard: what died in two years
Operator, Atlas, Agent Builder, IBM's ACP — flagship agent surfaces that died in months, and the three patterns in how they went.
2026-07-24
MCP vs A2A: which layer each protocol owns
They aren't competitors — they're adjacent layers. Where the seam sits, how both races ended, and when your architecture needs each one.
2026-07-24
Permission modes: graduated autonomy for AI agents
The four-rung autonomy ladder shipped agent products converged on, and the rules that make it earn trust instead of burning it.
2026-07-24
The approval gate: sign-off that stays meaningful
Why agent approval gates decay into click-throughs, and the five design moves that keep human sign-off worth something.
2026-07-24