Published: 2026-07-14

Nick Saraev's AI Agent Workflow: Linear + Webhooks + Fable 5

Chapters / key moments (click to jump — plays here on the page)

Nick Saraev walks through the AI-agent workflow his business actually runs day to day: a shared human-and-AI workspace in Linear where tasks tagged "AI ready" are picked up by an agent through a webhook. The agent — powered by Fable 5 — consults a knowledge base for context and preferences, does the work autonomously, and waits for human approval before any outward-facing action.

Source video

"Steal My Actual AI Agent Workflow (2027)" by Nick Saraev — Watch on YouTube →

Key Takeaways

Start with a shared human+AI workspace. The shape matters more than the tool. He uses Linear with statuses that denote pipeline position: inbox → next → doing → waiting → done, so any task's state (and how it was completed) is legible to both people and agents.
Hand tasks to agents, not just people. Write the task, describe what you want in plain language, tag it "AI ready," then drag it to "next" — which fires a request to a server where the agent lives.
The agent runs a real loop. His "Nick OS" run assesses the card, posts a plan, then executes — consulting a knowledge base for how prior tasks were done and his stated preferences before writing anything.
Fable 5 gets three inputs. Workspace context (files, other tasks, the content pipeline), a separate text knowledge base it can draw on, and the task itself.
Human-in-the-loop by design. For anything outward-facing — publishing, posting to a profile, touching the wider internet — the agent pauses for approval. You act as the project manager over a queue rather than babysitting each step.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Configuration shown

Task board statuses (Linear)

inbox → next → doing → waiting → done. Humans and agents share the board. A task tagged "AI ready" (also called agent ready) and moved to next is picked up by the agent; the agent moves it to doing, then to waiting when it needs the human.

Output sizes in his visual check

Thumbnails 1280 by 720; diagrams 1920 by 1080. Style he specifies: black ink on pure white, editorial look.

His scoring gate (as described)

Five questions each scored 0 to 2 (first-principles reasoning, expected-value discipline, whether his time was minimised, verified rather than merely plausible, a leverage check). Below a total of 7 the task fails and the agent keeps iterating until it passes.

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

1. A shared human and AI workspace
Any project tool will do; he uses Linear. The shape matters more than the product: one board, a queue of tasks, and statuses that show where each is.
2. Webhook to an agent
When an issue is tagged and moved to next, Linear sends a request (a webhook) to a server where his agent runs. The agent (Fable 5 in the demo) reads the workspace context, a knowledge base of his preferences and past work, and a set of credentials, posts a plan on the card, and starts working.
3. Approval stop
For anything involving publishing or the wider internet, the agent waits in the waiting status for his approval.
4. Low-friction capture
A keyboard hotkey and an iPhone action-button shortcut create a Linear task in a second; pressing twice can mark it as an agent task, once as a human task. Push notifications tell him when an agent is waiting on him.
5. Evals before it reaches him
Every output passes a checklist first: text renders cleanly with no garbled characters, matches his style, is legible at thumbnail size, is faithful to the brief and the right size, plus his tone of voice and no typical AI tells such as em dashes. If it fails, the agent retries until it fits.
6. Final QA is the human's job
He batches check-ins a few times a day and spends his time scoping tasks and reviewing results.

Gotchas

He says agent output is not incredible on its own: verify it, force ideation over a large space and then apply your own taste to pick winners.
In the demo the agent posted using his own Linear profile to avoid paying for another seat.
He says you can paste the video's transcript into Fable or GPT-5.6 and have it recreate a similar system, using roughly 25% to 50% of a session limit.
Evals cost tokens; he judges that worth it against the time spent fixing an average task by hand.
Promotional ending for his paid programme. His revenue and productivity claims are unverified.