Killed By Claude Report
← HomeCurrent verdict
Assessment
Guiding Hand’s core value proposition is: let an AI safely act on someone else’s computer.
Anthropic has already announced or signaled:
- stronger computer use,
- recurring agent task execution,
- risk classification before destructive actions,
- file and local environment access,
- and explicit plans for local support.
That is uncomfortably close.
Guiding Hand is selling a more operationally complete product for support teams:
- customer invite links,
- consent-first connection flow,
- active stop controls,
- approval checkpoints,
- API embedding into helpdesk workflows,
- recordings and audit trails.
Anthropic’s evidence shows capability expansion, but not a dedicated external-customer remote support product. Claude is becoming the brain. Guiding Hand is trying to be the execution and compliance wrapper around that brain.
That can still matter.
Biggest historical hit
The hardest punch here is "Introducing Sonnet 4.6" because it explicitly calls out improved computer use, agent planning, and stronger execution quality.
Guiding Hand depends on the idea that controlling a remote computer through natural-language tasks is specialized infrastructure.
Every time Anthropic improves the underlying computer-use model, more of Guiding Hand’s perceived magic collapses into the base model layer. If Claude gets good enough at reliable GUI execution, Guiding Hand risks being seen as a thin workflow shell unless it owns the trust, deployment, and support-specific control plane.
What still protects them
Guiding Hand still has a few real defenses, if they execute:
- Consent and trust UX
Their product is built around the fact that the computer belongs to a customer, not the operator. That changes everything: onboarding, permissions, visible stop controls, risky-action approvals, and session boundaries.
- External-machine orchestration
Anthropic can add local support, but that is not the same as making it easy for a support team to connect to thousands of non-technical customers’ laptops.
- API-native support embedding
If they become the standard infrastructure inside helpdesks, MSP tools, and service workflows, they can sit underneath whichever model wins.
- Auditability
Events, history, and replay are not glamorous, but enterprises buy them.
- Model-agnostic execution layer
The smartest move is to avoid pretending they are the intelligence layer. They should be the safest and easiest way to let any agent act on an end-user machine.
If they fail to deepen these advantages, Claude eventually eats the story.
Signals
Why this is in the blast radius
Sonnet 4.6 explicitly improves computer use, agent planning, and instruction following.
That attacks Guiding Hand at the capability layer: their product depends on an agent being able to reliably carry out natural-language actions on a computer. Better base-model computer control makes Guiding Hand less differentiated unless its moat is the session infrastructure, consent workflow, and enterprise controls.
Guiding Hand emphasizes approvals for risky actions and structured human oversight.
Anthropic publicly describes a classifier that reviews tool calls for destructive actions, allows safe ones, and blocks risky ones. That is not identical to Guiding Hand’s customer approval UX, but it overlaps with the same safety narrative and chips away at the uniqueness of their guardrail layer.
Anthropic says threads already run in the cloud and that local support is coming very soon.
Guiding Hand exists to let agents act on real end-user computers. If Claude gains native local-machine reach, the substitution risk rises sharply. The caveat is important: the tweet admits Claude cannot yet reach local files, tools, or internal networks at that moment, so this is a threat trajectory more than full present-day replacement.
Anthropic describes assigning tasks that Claude completes on your computer and handling recurring workflows like scanning email or pulling reports.
That overlaps with Guiding Hand’s broader promise of delegating desktop actions to an agent. It is less directly targeted at customer support remoting, but it normalizes Claude as a general-purpose computer operator.
Folder-level read, edit, and create access shows Anthropic extending Claude from chat into local execution contexts.
This does not equal remote support control of a customer desktop, so it does not replace Guiding Hand’s session product. But it weakens the argument that local machine task execution needs a separate vendor at all.
This announcement focuses on Claude as a productivity hub with third-party integrations and future agentic commerce.
That matters strategically because Anthropic is expanding Claude into a broader action surface. But it does not specifically show customer-computer remote control, support-session management, or consented external device access. So it is adjacent, not a direct hit.