All posts
5 min read

What AI agents still can't do (from people who ship one)

We run an AI agent platform, which makes this an odd post to write and exactly why it's worth reading. Every vendor tells you what agents can do. Here is the other list — the one we publish on our own pages, kept current because pretending otherwise costs more than admitting it.

They only reach what you connect

An agent has no ambient access to your life. It can search the web and read what it finds, but your inbox, calendar, repos and CRM are reachable only after you connect them — the connections you make are the entire perimeter. On Choreo that's 68 connectors plus any HTTP API you describe, and the flip side is a guarantee about credentials: a service you never linked is a service no agent of yours can act on as you. It can read the public web; it cannot log in anywhere you didn't connect. And every call to a linked service is logged with what it cost.

They can be wrong, and big builds take passes

A model can misread a page, mis-total a spreadsheet, or confidently summarize the wrong thing. Two structural facts make this permanent rather than a bug to be patched:

  • Context compaction is lossy. Long tasks outgrow the model's context window, and the summarization that keeps them going discards detail. We built a separate memory system precisely because context cannot be trusted to survive — and we say "lossy" in our own agents' instructions.
  • Sub-agent output is unverified evidence. When agents delegate, the parent is instructed to treat the worker's confident summary as a claim to check, not a fact to forward. That discipline exists because the failure it prevents is real.

Practically: a "build me an app" request produces a real first pass, and then it takes steering — a few rounds of "the header is wrong, the form should do X" — to get something you would actually ship. Anyone promising one perfect shot is describing a demo.

Some gates we won't open

A few limits are deliberate. Some are policy: agents don't give legal, medical, tax or financial advice, and nothing gets signed by software. And one set is enforced in code rather than promised in prose — when an agent runs on a schedule with nobody watching, the permission model hardens: anything configured to ask for approval is simply denied — and email sending, credential access, file deletion and computer control can never be pre-approved for unattended runs at all. Not "we advise against it." The grant option does not exist. That is the worked example of how we'd rather lose a feature checkbox than ship a 3am surprise — the scheduling post walks through it.

Why we publish this

Because the alternative is worse. Our own feature pages end with "what they can't do" sections; the trust page states where the security program stands rather than implying completion — including admitting that the native web tools inside Codex and Claude Code bypass our own capability gate, which is the kind of sentence a vendor only writes if the rest of the page is true too. A tool you'll hand real work to should be evaluable, and limits are half of any honest evaluation. The other half is trying it against something real: test both lists yourself — the free plan needs no card.

See it for yourself — free, $2 of credits once you confirm your email, no card.

Start free