Integration Dead-Ends

#concept

Summary

A distinct agent failure mode, separate from hallucination: the agent accepts a task and starts working against a capability the user doesn't actually have — a connector that isn't implemented, or one gated behind an account tier, paid seat, or corporate plan. The user only finds out after the time and tokens are already spent. larysa: "That's not hallucination exactly… it just first goes looking for what you don't have."

Current Understanding

The failure is one of optimistic capability assumption. The agent reasons about the ideal tool for the job rather than the tools available to this user in this account, and nothing in the loop checks entitlement before work begins. Cost is asymmetric: the discovery happens late (half an hour in, in Larysa's case), so the whole run is wasted rather than cheaply rejected.

Three concrete gates seen in the corpus:

No external API at all — Microsoft Teams. Asserted as an outright dead end for a chat-watching auto-task bot.
Paid seat required — Slack has first-class Claude support from Anthropic, but the seat wasn't granted; Figma's editable access needs a personal paid account.
Account type — the recommended path existed only for corporate, not private, accounts.

The corpus offers no fix, only a discipline: verify the integration and account requirements up front (connector exists? private vs. corporate? paid seat?) before letting the agent spend. Both participants explicitly left the general problem unsolved. Structurally this is the mirror image of leave-less-room-for-imagination — there the agent over-interprets an under-specified goal; here it over-assumes an unverified capability.

Note the enterprise angle: entitlement gates are precisely what a centrally-managed harness could encode once for everyone, rather than each user rediscovering them by burning tokens (enterprise-ai-reality).

Evidence

The Teams/ClickUp bot dead end, private-vs-corporate account gate, the "half an hour" cost, "I don't know how to fight this" — 2026-07-21-larysa-interview.
Slack/Figma paid-seat blockers from the same interview.

Concepts: leave-less-room-for-imagination, skills-as-memory (a skill can freeze "what we actually have access to"), context-as-scarce-resource, enterprise-ai-reality, evolution-of-agent-tooling
Entities: larysa, eugene, claude-code

Contradictions / Uncertainty

"Teams has no usable external API" is an assertion by the participants, not a demonstrated finding. Status: tentative.
Whether this is a harness bug (no entitlement pre-check) or a user-workflow gap is unresolved in the corpus.

Next Questions

Could a standing "available integrations" context doc or skill — an explicit capability inventory — pre-empt this cheaply?
Do agents fail this way on local capabilities too (missing binaries, no credentials), or is it specific to hosted connectors?

3.0 KiB Raw Permalink Blame History