A weekly publication by Daniel Ternyak

What it takes to build an agent into your product.

Teardowns of the agents Shopify, Notion, Linear, Intercom and others already shipped, read for the decisions they hand your security, design, platform, support and finance teams.

Latest teardown#

All 12 teardowns

Intercom FinLatestFin is paid per outcome, and Fin writes the definition of an outcome.Fin is sold per outcome instead of per seat. Here’s how it works, the decisions behind it, and what to check before you trust the number it bills you on.Price$0.99 per outcome, charged at most once per conversation.ModelFin Apex 1.0, trained in-house.Claimed results76% average resolution rate across 12,000+ customers, and about 2 million resolutions a week.Read the Intercom Fin teardown

What the product does that the model can’t#

One request, traced through the product’s own services. The model proposes; the product checks, asks and records.

Run the same request asProduct agentChatbot bolted on

The product’s own services

MLMayaPMMove the launch to Friday

Modeldecides

proposesLaunch Fri12 tasks +1 dayEmail customers Email: needs approval

Data modelget launch12 tasks, owners, announcement draft

Permissionsas Maya edit tasks email customers

Schedulingvalidate Fri Friday is valid 1 owner out

Launch screendraft~~Thu~~ Fri+12 tasksApprove Approved

Write APIcommit 13 saved · op_7f3asame API as the UI

Audit logreceipt #481Agent for Maya · 13 movedemail not sent

Historysnapshot Undo 13nothing was sent

Database 13 rows · op_7f3a

The trace, step by step:

Maya, a PM, asks: “Move the launch to Friday.”
The model decides on a plan: move the launch and its 12 tasks, and email customers.
Data model: the launch has 12 dependent tasks, their owners and an announcement draft.
Permissions, acting as Maya: editing tasks is allowed; emailing customers isn’t, so the email waits for approval.
Scheduling: Friday is a valid date; one task’s owner is out that day.
Launch screen: the change appears as a draft in the product’s own launch screen, and Maya approves it there.
Write API: the product saves the change through the same API its screens use, as one operation, op_7f3a.
Audit log: the agent, on behalf of Maya, moved the launch and 12 tasks; the email wasn’t sent.
Version history: undo restores all 13 items; nothing was sent.

The model proposes; the product enforces. The pattern behind each service:

The archive#

Every earlier teardown, each with its verdict and the decisions to steal.