AgentStatus
We tested every public MCP server we could find. Calling is solved. Delivery is not.
Read the reportResearch, engineering insights, and perspectives on AI agent reliability and infrastructure.
The live blog lists 47 posts.
MCP standardized how an agent calls a tool. It did not standardize how the agent is supposed to behave once it has one.
If MCP was the DoorDash to your product, WebMCP is the waiter who hands the menu to agents visiting your website.
Agents do not file tickets when they cannot use your site. They pick the next catalog. Build publishes an MCP from a website or dashboard, then we monitor it on the same tabs.
Live Traffic is the tool calls and jobs that ran for real use. MCP records host tool calls. Personal agents record calendar, mail, and booking jobs.
Personal agents do their work in Messages, on a phone you do not hold. We already checked tools inside Claude Code and Cursor. iMessage is the same floor, in the place those agents actually live.
A ChatGPT plugin is an HTML widget a tool can attach to its result. Agent Status opens that widget in a host sandbox and scores whether the host can show it. Tools still working is a separate question.
A person asks for a refund. The tool starts it. The money clears later. ChatGPT only hears that if it subscribed.
Spec 2026-07-28 ships today: initialize gone, sessions gone, _meta on every request. Why the migration window breaks agents quietly — and what dual-era independent third-party validation has to prove.
MCP sits between your agent and every tool it can call. When transports drift, tools vanish, or gateway connections silently fail, your dashboard stays green and your agent quietly lies.
The alert fired. The dispatcher logged it. No one received it. Six failure modes in the alert pipeline most teams never instrument.
Stuck workers, hanging streams, recursive planners, and silent background tasks. The most expensive agent failures happen in code paths your HTTP-based observability was never designed to watch.
Dashboards are green, users are having a bad time. That gap between what your observability tells you and what users actually experience is what user-side monitoring is built to solve.
An illustrative benchmark across eight major model providers from three regions. TTFB, p99 tails, geographic variance, and a heuristic for picking a provider.
A walkthrough of the monitoring setup we run on our own agent. Three incidents it caught in 90 days before any customer filed a ticket, and how to copy it.
We analyzed thousands of validation checks across our global network. Average agent uptime is 97.2%, but effective uptime drops to 84.8% when you apply semantic validation.
12% of agents that return HTTP 200 are functionally broken. The monitoring dashboard shows 100% uptime. Customer complaints eventually surface the issue.
Uptime measures whether your agent responds. It doesn't measure whether your agent works. For AI agents, that distinction is everything.
Tests verify that code works in a test environment. Production validation verifies that your agent works in production, right now, from real locations.
Most teams deploy, hope it works, wait for customer complaints, then firefight. There's a better way.
We're in the 'Kubernetes in 2015' moment. The technology works. The ecosystem is immature. This is what the next few years look like.
300,000+ custom agents deployed in 2025. $4.2B in agent infrastructure investment. The real opportunity isn't building agents, it's the infrastructure they need.
TTFB and latency sound similar but hide very different problems. Here's how to tell them apart, measure each, and fix the one costing you users.
The gap between 'works in dev' and 'works in production' is where most agent failures live. Eight failure modes and how to prevent each one.
Within 18 months, every company with production AI agents will have agent-specific monitoring. Not because they want to. Because they have to.
For decades, SLAs meant uptime. For AI agents, uptime is necessary but not sufficient. The next generation includes semantic correctness.
Most teams drastically underestimate both detection time and impact per hour. The ROI on monitoring is 10,400%. Even at 5x lower estimates, it's 2,000%.
APAC users experience 4x the latency. Agents are 7x more likely to fail for APAC users than US users. Your US-only monitoring sees nothing.
Any agent that can't answer 'What is 2+2?' is broken. Full stop. That's the insight evaluation prompts encode.
Modern applications rely on multiple agents collaborating. This creates a new challenge: how do you trust a system where no single component is fully in control?
Agent marketplaces are booming. Platforms listing thousands of agents. There's just one problem: how do you know any of them actually work?
Live Traffic is a tool that already ran. Discoverability is the moment before that, when Claude decides whether your MCP is even in the room. We open a fresh chat, with yours not installed, and ask.
Your MCP server worked in development. Cursor listed the tools. A call returned the right payload. The demo looked clean.
Here is the prediction: every team that ships an MCP server into Cursor, Claude, VS Code, ChatGPT, or an in-house host will run MCP-specific validation.
SLAs for MCP are about to change, and the teams still quoting "99.9% uptime" will be the last to notice.
Every team with an MCP server in production has a version of this thread, usually after the person has already worked around you:
There is a failure mode sitting under every MCP you ship, and most stacks never name it.
It is the first question everyone asks. It is the wrong question.
"We ping the MCP URL in CI. Isn't that enough?"
If you run an MCP server in production, you already know the uncomfortable version of green: your process is up, discovery returns a list, the gateway logged 200, and a person in Claude still cannot finish fetch. Or Chat
Every dashboard you have was designed for a world where a green URL means the work finished.
Teams no longer ship a server for a single client. The same URL is opened by Cursor, Claude, VS Code, and ChatGPT. Each host initializes on its own terms. Each host is shown a catalog. Each host walks search, then fetch,
Directories, storefronts, partner catalogs, internal registries. Publishers paste a URL. Buyers browse, install, and open a host. The listing shows a name, a description, and often a green reachability mark.
Here is a question a URL check cannot answer:
A tool can sit in the catalog, look healthy to every URL check, and still fail the job. Hosts pick another tool. Support guesses. The bill is the time you spent looking at the wrong layer.
MCP is not a sidecar to your API. It is the product surface named hosts actually call. The host is the runtime. Validation has to sit outside both.
Inside a host, your server is not a brand. It is a set of tools next to other tools. The ones that finish the work get called again. The ones that only list get skipped.
Most MCP teams still ship, hope the URL stays up, and wait for someone to describe a foggy agent failure. The proactive version is a pin: host, step, first break after you shipped.