AI Engineer · Published 2026-10-05

Lifestyles of the AI-Native — Nick Nisi & Zack Proser, WorkOS

privateReview: unreviewedConfidence: high
agentsvoice-codingloopsverification-gatesfleet-managementagent-loopsadversarial-reviewscheduled-tasksmulti-modeldelegation-protocol
Open on YouTube ↗

Summary

Overview

  • Speaker: Nick Nisi & Zack Proser
  • Channel: AI Engineer
  • Main topic: Engineering workflows and fleet management for AI-native developers using Claude, voice coding, loops, and verification gates.
  • Purpose: To demonstrate advanced, practical workflows for AI-native software engineering, scaling from individual prompt-response loops to autonomous multi-agent fleet operation. Nick Nisi and Zack Proser from WorkOS present a workshop on operating a fleet of AI coding agents rather than babysitting a single session. They cover four core blocks: voice coding for high-bandwidth thought-to-code iteration, loops and goals for unattended execution, verification gates for robust error-free PR generation, and scheduled tasks for recurring maintenance.

Topic Map

Introduction and the AI-Native Premise

  • Explanation: Shifting the engineer's mindset from babysitting a single chat session to operating a fleet of asynchronous AI coding agents.
  • Key claims:
    • Most engineers babysit a single session, which is obsolete.
    • Operators should run a fleet of agents.
    • AI enables managing dozens of repositories across multiple languages with tiny teams.
  • Examples:
    • Two engineers at WorkOS managing 25+ repos across 8 languages using AI.
  • Terminology:
    • fleet
    • applied AI
    • WorkOS
  • Why it matters: Establishes the core philosophy of scaling engineering output through multi-agent orchestration.

Voice Coding

  • Explanation: Leveraging voice-to-text dictation tools (like Handy or WhisperFlow) to increase input bandwidth and code at the speed of thought.
  • Key claims:
    • Speaking is about 2x faster than typing (~184 wpm vs ~90 wpm).
    • Focus on stating the outcome rather than writing individual keystrokes.
    • Local on-device dictation tools provide privacy and eliminate subscription costs.
  • Examples:
    • Using Handy or WhisperFlow for rapid ideation and dictating complex refactors across files.
  • Terminology:
    • voice coding
    • dictation
    • bandwidth gap
    • Handy
  • Why it matters: Removes the typing bottleneck and matches input speed to mental ideation.

Loops and Goals

  • Explanation: Moving from manual prompt iteration to goal-driven and recurring loops that run unattended until conditions are met.
  • Key claims:
    • An agent is a model with tools in a loop (think, act, observe, decide).
    • Goals are bounded conditions (run until tests pass); loops are recurring tasks.
    • Done is a checklist, not a vibe.
  • Examples:
    • Running /goal bun playground/goals/check.ts to iterate until all checks pass autonomously.
  • Terminology:
    • /goal
    • /loop
    • state machine
    • verification gates
  • Why it matters: Enables engineers to hand over whole jobs and walk away, shifting from babysitter to operator.

Verification Gates

  • Explanation: Implementing automated checks, hooks, and multi-model adversarial reviews to ensure agents never hand over broken code.
  • Key claims:
    • Trust the gates, not the model's confidence.
    • A hook that lints, typechecks, and tests on every edit prevents broken code from reaching PR.
    • Fan out to a second model for adversarial review to catch blind spots.
  • Examples:
    • Using /codecs/adversarial-review to have Codex review Claude's PR diffs.
  • Terminology:
    • verification gates
    • hooks
    • adversarial review
    • linting
  • Why it matters: Solves the trust problem in AI code generation by enforcing objective programmatic quality standards.

Scheduled Tasks

  • Explanation: Automating recurring maintenance, dependency bumps, status reports, and cron jobs using persistent cloud routines.
  • Key claims:
    • Schedule handles the toil that comes back every week.
    • Persistent cloud routines run even when you are offline.
    • Integrate scheduled tasks with Slack and GitHub to automate operational overhead.
  • Examples:
    • /schedule every weekday 9am run bun playground/scheduled/report.ts
    • Weekly token usage and PR summary dashboards.
  • Terminology:
    • /schedule
    • cron jobs
    • toil
    • fleet management
  • Why it matters: Automates routine developer toil, freeing up human engineers for high-leverage creative work.

Key Points

Operate a fleet, don't babysit sessions

  • Explanation: Engineers should manage multiple asynchronous AI agents across projects rather than interacting with a single chat prompt sequentially.
  • Evidence: Two engineers managing over 25 repositories across 8 languages.
  • Practical implication: Multiplies individual developer productivity significantly.

Voice is high-bandwidth input

  • Explanation: Dictation tools allow developers to express architecture and intent at nearly 200 words per minute.
  • Evidence: Speaking speeds (~184+ wpm) drastically outpace typing speeds (~90 wpm).
  • Practical implication: Reduces physical friction and speeds up initial ideation and prompting.

Make 'done' concrete with verification gates

  • Explanation: AI models declare victory early and produce confidence without correctness. Programmatic gates enforce objective completion.
  • Evidence: Checklists requiring passing tests, linting, and typechecks before merging.
  • Practical implication: Eliminates shipping broken code generated by hallucinating agents.

Frameworks, Models & Processes

The Agentic Loop

  • How it works: The model cycles through thinking, acting with tools, observing results, and deciding the next step until a goal condition is satisfied.
  • Components:
    • Think
    • Act
    • Observe
    • Decide
  • When to use: For any automated task or multi-step coding workflow.

The Four-Box Goal Framework

  • How it works: A strict checklist for agent task execution where no box is optional.
  • Components:
    • Reproduce the bug with a failing test
    • Implement the smallest fix
    • Lint, typecheck, and ensure tests pass
    • Open a PR with the plan in the body
  • When to use: When delegating bug fixes or feature development to an autonomous agent.

Examples & Case Studies

Using voice dictation in Claude Code to set up workshop tools and request repo restructuring.

  • Illustrates: Voice coding and outcome-based prompting.
  • Lesson: Say the outcome rather than typing keystrokes to move at the speed of thought.

Running /goal with a check script to iterate on code until all tests pass without human intervention.

  • Illustrates: Autonomous goal-driven execution.
  • Lesson: Let agents loop and self-correct against programmatic judges.

Actionable Takeaways

  • Immediate:
    • Adopt a voice dictation tool for coding and planning.
    • Define explicit verification gates (lint, typecheck, test) for AI outputs.
    • Use /goal and /loop primitives to delegate multi-step tasks.
  • Strategic:
    • Shift engineering roles from writing code to operating and supervising AI agent fleets.
    • Automate recurring developer toil using scheduled agent routines.
    • Implement multi-model adversarial reviews to catch agent blind spots.
  • Questions to investigate:
    • How do we manage token costs and burst limits when running 10+ agents simultaneously?
    • What are the best practices for secure secret management in unattended cloud agent loops?

Claims Worth Verifying

  • Speaking is about 2x faster than typing for coding prompts. (metric)
  • Verification gates prevent agents from handing over broken code. (technical principle)

Notable Quotes

"Most engineers babysit a single session. Operators run a fleet." "Done is a checklist, not a vibe." "Trust the gates — not the model's confidence."

Compressed Summary

  • Shift from babysitting single sessions to operating agent fleets
  • Use voice coding for high-bandwidth ideation and prompting
  • Employ goal and loop primitives for unattended execution
  • Enforce verification gates and adversarial reviews to guarantee code quality
  • Keywords: agents, voice-coding, loops, verification-gates, fleet-management
  • Core insight: AI-native engineering requires transitioning from manual typing and single-session babysitting to operating verified, automated fleets of goal-driven coding agents.

Core insights

7
Architecturemedium noveltystrong evidence

Trust in agent output should be relocated from model self-reported confidence to programmatic verification gates (lint, typecheck, test hooks run on every edit) that physically block broken code from reaching a PR.

Why it matters

Sets the trust boundary outside the model, making correctness a property of the harness rather than the generator and eliminating the dominant failure mode of agents hallucinating success.

Generalization

Any generative pipeline can replace 'ask the model if it's done' with deterministic gates that gate promotion to the next stage.

Trust the gates, not the model's confidence.
Open source video
A hook that lints, typechecks, and tests on every edit prevents broken code from reaching PR.
Open source video
AI models declare victory early and produce confidence without correctness.
Open source video
Mechanismmedium noveltystrong evidence

Closing the loop with a programmatic judge: a goal is a bounded termination condition (e.g. run until check script passes) distinct from a recurring loop, letting the agent self-correct against an external check rather than a human in the loop.

Why it matters

Turns the agent loop from a chat turn into a state machine with an objective stop condition, which is what makes unattended execution safe enough to walk away from.

Generalization

Separating bounded goal conditions from recurring schedules is a reusable control-plane primitive for any autonomous task runner.

An agent is a model with tools in a loop (think, act, observe, decide).
Open source video
Goals are bounded conditions (run until tests pass); loops are recurring tasks.
Open source video
Let agents loop and self-correct against programmatic judges.
Open source video
Practicehigh noveltymoderate evidence

Cross-model adversarial review: fan the PR diff out to a different model than the one that produced it (Codex reviewing Claude's diffs) to catch blind spots the generator is structurally unable to see.

Why it matters

Suggests single-model self-review is an insufficient safeguard and that verification should be a separate, ideally vendor-independent service.

Generalization

Independent adversarial verifiers with different training/weights outperform self-critique for reducing correlated errors.

Fan out to a second model for adversarial review to catch blind spots.
Open source video
Using /codecs/adversarial-review to have Codex review Claude's PR diffs.
Open source video
Mechanismmedium noveltystrong evidence

A strict four-box delegation protocol where no step is optional: reproduce the bug with a failing test, implement the smallest fix, lint/typecheck/tests pass, open a PR with the plan in the body.

Why it matters

Encodes 'done' as a fixed checklist rather than a judgment call, giving a reproducible contract for delegating bug fixes to autonomous agents.

Generalization

Encoding completion as an ordered, non-skippable artifact-producing checklist is portable to any delegated task, not just code.

Done is a checklist, not a vibe.
Open source video
Reproduce the bug with a failing test
Open source video
Implement the smallest fix
Open source video
Open a PR with the plan in the body
Open source video
Mental Modelmedium noveltymoderate evidence

Shift from babysitting one interactive session to operating a fleet of asynchronous agents, enabling tiny teams to manage many repos and languages.

Why it matters

Reframes the engineer's role from synchronous prompter to supervisor/scheduler, changing what tooling and observability matter (parallel cost, secrets, concurrency).

Generalization

As agent reliability rises, value migrates from interaction skill to orchestration, supervision, and resource governance.

Most engineers babysit a single session, which is obsolete.
Open source video
Operators should run a fleet of agents.
Open source video
Two engineers at WorkOS managing 25+ repos across 8 languages using AI.
Open source video
Practicemedium noveltymoderate evidence

Voice dictation as high-bandwidth, outcome-oriented input: ~2x speaking-to-typing speed ratio, with local on-device tools chosen for privacy and no subscription.

Why it matters

Addresses the human input bottleneck to agentic iteration and shifts prompting toward stating outcomes rather than keystrokes.

Generalization

Whenever human intent generation exceeds the input channel's bandwidth, upgrading the channel is a leverage point.

Speaking is about 2x faster than typing (~184 wpm vs ~90 wpm).
Open source video
Focus on stating the outcome rather than writing individual keystrokes.
Open source video
Local on-device dictation tools provide privacy and eliminate subscription costs.
Open source video
Practicemedium noveltymoderate evidence

Persistent cloud routines for recurring toil (scheduled tasks, cron jobs, dependency bumps, weekly reports) that run even when the developer is offline and integrate with Slack/GitHub.

Why it matters

Extends agent operation from on-demand to always-on infrastructure, which introduces new cost, secrets, and scheduling concerns.

Generalization

Persistent, event/time-triggered agent routines are infrastructure, not sessions, and inherit infrastructure's operational requirements.

Persistent cloud routines run even when you are offline.
Open source video
/schedule every weekday 9am run bun playground/scheduled/report.ts
Open source video
Schedule handles the toil that comes back every week.
Open source video

Deep dives

4

Cross-model adversarial review as a decorrelated verifier

Research question

Does a different-model reviewer catch defects the generating model's self-review misses at a rate that justifies the added latency and spend, or does it mostly add correlated noise?

Why

If verification is only credible when the verifier's errors are decorrelated from the generator's, then 'review by a second vendor' becomes an architectural requirement rather than a nice-to-have — and the cost model of every agent pipeline changes.

Fan out to a second model for adversarial review to catch blind spots.
Open source video
Using /codecs/adversarial-review to have Codex review Claude's PR diffs.
Open source video
Source video

Verification gates as the real trust boundary

Research question

What proportion of agent-declared completions fail independent lint/typecheck/test gates, and does gate enforcement measurably lower human rework per delegated task?

Why

Relocating trust from model confidence to deterministic harness checks is the load-bearing claim behind unattended execution; without a measured failure rate it is an assertion, not an engineering result.

Trust the gates, not the model's confidence.
Open source video
A hook that lints, typechecks, and tests on every edit prevents broken code from reaching PR.
Open source video
AI models declare victory early and produce confidence without correctness.
Open source video
Source video

Cost, concurrency and secret governance for always-on agent fleets

Research question

What budget, rate-shaping and credential-scoping controls are required before N concurrent agents can run unattended without runaway spend or over-privileged credentials?

Why

Fleet operation converts a productivity gain into an infrastructure problem: parallel token spend, burst limits, and long-lived secrets in offline routines are the failure modes that decide whether fleets are viable in production.

How do we manage token costs and burst limits when running 10+ agents simultaneously?
Open source video
What are the best practices for secure secret management in unattended cloud ag
Open source video
Persistent cloud routines run even when you are offline.
Open source video
Source video

Bounded goals vs recurring loops as a control-plane primitive

Research question

Can a single runtime that distinguishes bounded termination conditions from recurring schedules make unattended execution safe and auditable across heterogeneous task types?

Why

Separating 'run until the check script passes' from 'run every weekday at 9am' is what turns a chat turn into a state machine with an objective stop condition — the precondition for letting engineers walk away.

An agent is a model with tools in a loop (think, act, observe, decide).
Open source video
Goals are bounded conditions (run until tests pass); loops are recurring tasks.
Open source video
Let agents loop and self-correct against programmatic judges.
Open source video
Source video

Article ideas

4

Stop Asking the Model If It's Done

Correctness in agentic pipelines must be a property of the harness, not the generator: any pipeline that accepts the model's self-reported success as a promotion signal is one hallucinated victory away from shipping broken artifacts.

Angle

Argue for moving the trust boundary outside the model and show the concrete shape of that move — hooks that lint, typecheck and test on every edit, plus PRs that cannot open until gates pass.

Source video

Your Agent Fleet Is Infrastructure, and It Has a Bill

Once agents run on schedules while the developer is offline, they inherit every requirement of always-on infrastructure — budgets, burst limits, scoped short-lived credentials and per-agent accounting — and teams that treat them as sessions will discover this through an incident.

Angle

Frame the mental-model shift from babysitting a session to operating a fleet, then enumerate the operational surface the shift silently creates.

Source video

Self-Review Is Structurally Insufficient

A model reviewing its own diff cannot reliably catch the errors it is predisposed to make; decorrelated verification requires different weights, which makes the adversarial reviewer a separate service with its own cost and vendor coupling rather than a prompt.

Angle

Contrast same-model self-critique with a cross-vendor adversarial pass over the same PR diff, and be explicit about the added latency, spend and dependency.

Source video

Done Is a Checklist, Not a Vibe

Delegation only scales when completion is encoded as an ordered, non-skippable set of artifact-producing steps; reproducing the bug with a failing test before implementing the smallest fix is what makes autonomous bug-fixing auditable instead of merely plausible.

Angle

Take the four-box protocol apart and show why each step is load-bearing and what breaks when a step is skipped.

Source video

Project ideas

4

Gatehouse Completion Auditor

gatehouse

A large fraction of agent-declared task completions fail deterministic lint/typecheck/test gates, and enforcing those gates as non-bypassable pre-PR hooks reduces human rework per delegated task compared to diff review alone.

Proof of concept

Wrap an agent harness so every edit triggers a lint/typecheck/test hook, record every 'done' declaration, and measure gate-failure rate and loop iterations-to-pass against a control group without gates.

Measurement

Percentage of declared completions that fail gates; median loop iterations to first passing state; human rework minutes per delegated task with and without gates.

Source video

Decorrelated Reviewer Benchmark

beyond-evals

A second model from a different vendor catches seeded and real defects at a higher rate than the generating model's self-review, with a measurable but bounded incremental cost per PR.

Proof of concept

Build a diff-review rig that runs both self-review and cross-model adversarial review over a corpus of PR diffs with seeded defects, scoring catch rate, false-positive rate, latency and token spend.

Measurement

Defect catch rate and false-positive rate for self-review vs cross-model review; incremental tokens, latency and dollar cost per review.

Source video

Fleet Governor

new

Per-agent budgets, concurrency ceilings and short-lived brokered credentials make 10+ concurrent unattended agents economically and operationally safe without reducing completed-task throughput.

Proof of concept

A control plane that schedules recurring agent routines, assigns each a token/spend budget and a scoped credential lease, and hard-stops runs that exceed either, while logging per-agent cost and burst behavior.

Measurement

Cost per completed task at 1 vs 10+ concurrent agents; incidence of budget or burst-limit breaches; number of long-lived secrets eliminated from agent environments.

Source video

Input Bandwidth Probe

movement-lab

Voice dictation meaningfully raises effective specification throughput for outcome-oriented prompting in real refactoring tasks, but the gain is smaller than the naive ~184 vs ~90 wpm ratio suggests.

Proof of concept

Instrument paired coding sessions — dictation vs typing — on the same multi-file refactor tasks, capturing time-to-specification, correction rate and resulting artifact quality.

Measurement

Effective words-per-minute to a usable specification, dictation correction/edit rate, and task completion time for equivalent refactors.

Source video

Architectural implications

3

Verification is moved out of the model into hooks that fire on every edit.

Before

Trust the model's stated confidence and review output at the end.

After

Deterministic lint/typecheck/test gates intercept each edit before it can reach a PR.

Consequence

Correctness becomes a harness responsibility; models can be swapped without renegotiating trust.

Source video

Verification is decomposed into a separate cross-model service.

Before

One model generates and self-reviews its own output.

After

A different model adversarially reviews the generator's diffs.

Consequence

Introduces a new component, additional inference spend, and a dependency on a second vendor, in exchange for decorrelated error catching.

Source video

Execution is modeled as bounded goals plus recurring loops rather than chat turns.

Before

Human-driven prompt-response iteration.

After

A state machine with objective termination conditions and scheduled recurrence.

Consequence

Unattended execution becomes feasible but pushes requirements onto cost controls, secret management, and burst limits.

Source video

Tradeoffs and failure modes

4

Adversarial cross-model review

Benefit

Catches blind spots the generating model cannot see.

Cost or risk

Adds a second model's latency and spend to every review, and creates cross-vendor coupling.

Fan out to a second model for adversarial review to catch blind spots.
Open source video
Source video

Unattended fleet execution

Benefit

Enables handing over whole jobs and walking away.

Cost or risk

Token cost and burst-limit management when running many agents concurrently is unresolved.

How do we manage token costs and burst limits when running 10+ agents simultaneously?
Open source video
Source video

Persistent cloud routines

Benefit

Runs recurring maintenance even when the developer is offline.

Cost or risk

Raises secret-management exposure for unattended cloud agents.

What are the best practices for secure secret management in unattended cloud ag
Open source video
Source video

Voice coding

Benefit

~2x input bandwidth over typing.

Cost or risk

Requires adopting dictation tooling and favors local on-device tools to preserve privacy.

Local on-device dictation tools provide privacy and eliminate subscription costs.
Open source video
Source video

Open questions

4

How do we manage token costs and burst limits when running 10+ agents simultaneously?

Why unresolved

The summary raises it as an investigating question with no mechanism or answer given.

Research direction

Build cost/concurrency governance for fleets: budgets, rate shaping, and per-agent accounting.

Source video

What are the best practices for secure secret management in unattended cloud agents?

Why unresolved

Stated as an open question; no solution or pattern is provided.

Research direction

Evaluate scoped, short-lived credentials and secret-brokering for always-on agent routines.

Source video

Does a second, different model reliably catch errors the generator misses, or just add correlated noise?

Why unresolved

Adversarial review is demonstrated only as an example, with no measured error-catch rate.

Research direction

Benchmark cross-model adversarial review vs self-review on seeded-defect diffs.

Source video

Do verification gates plus bounded goal loops actually reduce human intervention rate in practice?

Why unresolved

Claimed as a workflow benefit but no measurement of intervention or rework rate is reported.

Research direction

Instrument fleets for gate-failure rate, loop iterations to pass, and human rework per delegated task.

Source video

Key claims

7
comparativeVerification needed

Speaking is about twice as fast as typing (~184 wpm vs ~90 wpm).

Evidence

Speaking is about 2x faster than typing (~184 wpm vs ~90 wpm).

Question

Do measured dictation throughputs in real coding workflows match ~184 wpm effective input?

Source video
causalVerification needed

AI models declare victory early and produce confidence without correctness, so programmatic gates are required.

Evidence

AI models declare victory early and produce confidence without correctness. Programmatic gates enforce objective completion.

Question

What proportion of agent-declared completions fail independent lint/typecheck/test gates?

Source video
factualVerification not requested

An agent is a model with tools in a loop (think, act, observe, decide).

Evidence

An agent is a model with tools in a loop (think, act, observe, decide).

Source video
factualVerification needed

Two engineers at WorkOS manage 25+ repos across 8 languages using AI.

Evidence

Two engineers at WorkOS managing 25+ repos across 8 languages using AI.

Question

Is the 25-repo/8-language coverage sustained by AI assistance or would it hold without it?

Source video
factualVerification not requested

Persistent cloud routines run even when the developer is offline.

Evidence

Persistent cloud routines run even when you are offline.

Source video
opinionVerification needed

Babysitting a single session is obsolete; operators should run a fleet of agents.

Evidence

Most engineers babysit a single session, which is obsolete.

Question

For which task classes does fleet operation outperform single-session interaction?

Source video
causalVerification needed

Inspecting PR diffs with a second model catches blind spots the generating model misses.

Evidence

Fan out to a second model for adversarial review to catch blind spots.

Question

What is the defect catch rate of a second model over self-review on real PR diffs?

Source video

Connections

5