Context ยท behind the scenes

How this got made

OpenAI asked me for one asset an enterprise developer would share internally to get their team excited about Codex, plus a short note on the thinking. This page is the note. It's also the director's commentary, with every easter egg explained.

Three pages make up the submission. My suggested order: two minutes in the review app, ninety seconds of infomercial, then back here for the why. The hub links all of them.

The note the brief asks for

Who it's for, why this workflow, why this format, and what's made up.

Audience

A tool only sticks when three people say yes.

  • The sponsorAn engineering leader worried about losing engineers
  • The championThe engineer bringing Codex to the team
  • The skepticThe engineers who review the agents' work

The worked example is Electric Sheep Energy, a fictional 1,000-person company that does diligence and procurement for clean-energy projects and data centers. Its 300 engineers have Codex through ChatGPT Enterprise. Most of them use it for autocomplete.

Why this workflow

Maintenance and inbound bugs eat the week. That's a retention risk. Engineers in Chainguard's survey:

16%of their week goes to building features
93%say features are the most rewarding work
2 in 3tech leaders worry about keeping talent

Handing the chores to agents is the obvious move. Then the skeptic asks who reviews all those PRs, and he has a point. Faros found AI-heavy teams get:

98%more PRs
91%longer reviews

So the human's job shrinks to one thing, the review queue. Every replicant PR gets a Codex review before a person sees it.

Why this format

Engineers mock marketing. They also share anything funny in #eng-general within about four minutes. So I made three pieces:

  • An interactive app engineers can poke at, where the workflow runs in front of them.
  • This context page, for anyone who wants the reasoning and the receipts.
  • An infomercial. If the marketing filter is going to fire anyway, I'd rather it fire at something worth forwarding.

Assumptions and simulated elements

  • Everything is fictional: the company, its rival, the people, the code and the numbers, including the rival's eNPS jump and the pilot results.
  • The GitHub sync in the review app is simulated. So is Approve. Nothing runs for real and nothing merges anywhere.
  • Codex's behavior is modeled on documented features. The Codex section spells out what's real.
  • The statistics are real and linked. The culture references are there for fun.

How I approached it

The calls I made along the way, and why.

  1. Enterprise, already paid for

    Electric Sheep already has ChatGPT Enterprise with Codex. Usage is low. That's the realistic enterprise problem: the budget fight is over, and the hard part is adoption and trust.

  2. Point of view

    The asset speaks in OpenAI's voice, so it reads like something OpenAI would publish and the champion can forward it without rewriting a word. This page is the one place I talk as myself.

  3. A company with something to lose

    Electric Sheep Energy: 1,000 people, about 300 engineers, checking whether clean-energy projects and data centers are worth funding. I wanted a business where one decimal point could cost a customer real money.

    Then I gave it a rival. Nexus Diligence quietly handed its engineers agents, its engineering eNPS went from +12 to +48 in a quarter, and nine Electric Sheep engineers left for it. The retention case makes itself.

  4. The cast

    Dana Holden is the champion. Leon Park is the brilliant, prickly reviewer, and also Dana's ex, who ghosted her. Marcus Gaff runs the next team and wants her job. Roy Batista is the VP who needs a plan by Thursday. Every adoption story has these four people. I just gave them names.

  5. The bug is human-made

    Dana typed the decimal point wrong, by hand, at 11pm. Codex review catches it. Codex never causes it. If an agent caused the bug, the whole story turns into a debate about trusting AI. A person's typo keeps it about review, which every engineer already cares about.

  6. The story style

    A plain, slightly uneasy narrator borrowed from Shirley Jackson's "The Lottery," echoed in a Monday chore draw nobody questions. Seven acts titled after Joseph Campbell's hero's journey. The tech shows up in one diff. Everything else is people.

  7. How the idea grew

    It started as a crew of maintenance replicants: dependency upgrades, flaky tests, security patches and deprecation warnings. Then I looked at where an engineer's interruptions really come from. Support tickets. A customer success manager's call recordings. Sales demo feedback. CI alerts. So the idea grew into this: every inbound request becomes a PR, and you only review. Maintenance chores are still in there, as one source among several.

What Codex actually does here

Engineers will fact-check a joke. Here's what's real, from OpenAI's current Codex docs.

"Runs on a schedule"
Codex scheduled tasks run on a recurring cadence. You pick it.
"The replicants build fixes"
Codex Cloud tasks work in the background and can open pull requests.
"Codex reviews every PR"
Automatic code review can be turned on for a repo. Commenting @codex review on a pull request asks for one by hand.
"Knows your rules"
Codex reads AGENTS.md before it works. Code review follows the rules in a "Code Review Rules" section of that file.
"A person approves"
Codex posts its findings as a review. Approving and merging stay with the team.
"Included with Enterprise"
Codex is included with ChatGPT Enterprise, and admins turn it on.
"Learns from your reviews"
When you request changes, your reason becomes a line in AGENTS.md, and Codex reads it on the next task. Codex doesn't train on your reviews. The team writes the rule and Codex follows it.
"Turns tickets into PRs"
Routing tickets, call notes and alerts into Codex tasks is the team's own plumbing. The review app shows that plumbing as if it's already built.

The story

The cast came from a short story I wrote first. It's seven acts, about four minutes, and entirely optional.

Read the story behind it 7 acts

TL;DR: Electric Sheep is losing engineers to a rival whose engineers don't do maintenance anymore. Dana wants to give every engineer an army of Codex replicants for the chores. Leon, the brilliant, arrogant engineer who ghosted her two years ago, thinks it will bury him in PRs, and Gaff, who wants her job, waits for her to fail. When a mistake Dana made by hand slips past Leon, she owns it and asks him for help. The replicants' reviewer catches what he didn't. The harder fix is between the two of them.

The characters
CharacterWho they are
Dana HoldenSenior engineer. Wants every engineer to have their own replicants.
Leon ParkOne of the best engineers in the company, and he knows it. Dated Dana, then stopped answering her texts.
Marcus GaffRuns the team next to Dana's. Wants her job.
Roy BatistaVP of Engineering. Losing engineers, and running out of time.

IThe ordinary world

Every Monday at Electric Sheep Energy, the engineers drew their chores. The tickets sat in a shared queue: dependency upgrades, flaky tests, security patches, the billing module nobody had touched since 2014. Each engineer took three. Nobody remembered who had started it, but everyone agreed it was fair.

Leon Park drew his without looking. He was the smartest engineer in the building and made sure you knew it. He answered questions with questions, and he sighed before he spoke in meetings.

Dana Holden sat two rows over. Two years ago, she and Leon had dated for five months. Then one week he stopped answering her texts. Her last one was still on her phone: "Are we okay?" He never replied.

IIThe call

In the spring, people started leaving. Nine engineers went to Nexus Diligence in one quarter. Their exit interviews said the same thing in different words: "I spent my last year upgrading dependencies." At Nexus, people said, the engineers didn't draw chores anymore. Agents did them.

Roy Batista, the VP of Engineering, called Dana into his office. "Give me a plan by Thursday," he said. "Before I lose anyone else." He asked how it was going every morning after that.

Dana proposed a team of Codex replicants for every engineer: one for upgrades, one for flaky tests, one for security patches. Codex would review everything the replicants opened, and a person would approve it. At the planning meeting, Leon leaned against the wall. "So instead of three chores a week, I review fifty robot PRs," he said. People laughed.

Marcus Gaff, who ran the team next to hers and wanted her job, smiled kindly. "Let her try," he said, loud enough for Roy to hear.

IIITrials

Dana started with one team and one repo. The replicants opened their first pull requests on Tuesday: small, tidy, each with a note explaining what might break. By Friday, the Platform team had cleared a backlog of upgrades older than most projects in the CAISO interconnection queue. Nobody drew chores that Monday. Some people seemed relieved. Some seemed uneasy, as if something had been taken from them.

IVThe ordeal

The replicants weren't the problem. Dana was. On the last Friday of the month, she pushed a fix she had typed by hand at eleven the night before. It needed Leon's approval. At 4:58, it came: "LGTM ๐Ÿ‘." He hadn't opened it. He never opened anything she sent.

On Monday, a customer called. Electric Sheep's report said their solar farm would lose half its power every year. Dana had put a decimal point in the wrong place.

By lunch, Gaff had a meeting with Roy. He brought one slide: "This is what happens when we let robots do the work." Roy paused the pilot and told Dana they would talk Thursday.

VThe dark night

Dana could have let the agents take the blame; Gaff already had. She could have blamed Leon; everyone had seen the timestamp. Instead she wrote to Roy: "The mistake was mine. I typed it, I didn't check it, and I sent it to someone I knew wouldn't read it."

Then she went to find Leon. He was still at his desk at eleven.

"You didn't read it," she said.

"You didn't check it."

"I know. I told Roy."

That made him look up.

"Why didn't you read it?" she asked.

He was quiet for a long time. "Same reason I didn't answer your text," he said. "It was easier not to."

VIThe Voight-Kampff

They didn't fix that, not that night. They fixed the other thing. They ran Dana's change through the same Codex review the replicants used, as if no one had ever approved it. It found the decimal point in four minutes.

"It read it," Leon said. "I didn't." It was the closest thing to an apology she had ever heard from him.

VIIReturn with the elixir

On Thursday, Dana showed Roy her note, the mistake, the four minutes, and the backlog the replicants had cleared. Gaff brought his slide again. Roy didn't look at it. He wrote to every manager, "We're doing this," and attached Dana's plan.

The next Monday, there was no chore draw. A few of the older engineers lingered by the queue anyway, as if waiting for their names. Dana used the hour to catch up on The Accidental Engineer, her favorite newsletter, for the first time in months. Leon read every change Dana sent. He argued about one variable name for twenty minutes.

That night her phone lit up. It was a reply, two years late, under "Are we okay?"

"No. That's on me. Coffee Monday?"

She looked at it for a while. Then she typed, "LGTM ๐Ÿ‘."

Easter eggs, behind the scenes

Some are for film fans. Some are for anyone who has been on call. Spoilers ahead, obviously.

In the story and the company

The references hiding in plain sight.

  • Philip K. Dick's Do Androids Dream of Electric Sheep?

    Where it hides
    The company name
    Who catches it
    Readers and sci-fi fans
  • Solar grazing

    Where it hides
    Also the company name: real solar farms use sheep to mow
    Who catches it
    Anyone in solar
  • Blade Runner's Voight-Kampff test

    Where it hides
    Code review as the test of what's real
    Who catches it
    Film fans
  • Holden, Roy, Leon, Gaff

    Where it hides
    Character names from Blade Runner
    Who catches it
    Deep-cut film fans
  • "LGTM ๐Ÿ‘"

    Where it hides
    The rubber-stamp review, and Dana's last word to Leon
    Who catches it
    Every engineer
  • Bikeshedding

    Where it hides
    Leon's 20-minute fight over a variable name
    Who catches it
    Every engineer
  • Being left on read

    Where it hides
    "Are we okay?", unanswered for two years
    Who catches it
    Anyone who has ever dated
  • CAISO interconnection queue

    Where it hides
    How old the upgrade backlog was
    Who catches it
    Solar and energy people
  • Joseph Campbell

    Where it hides
    The seven act titles
    Who catches it
    Storytellers
  • "How's it going?"

    Where it hides
    Roy's question, every morning
    Who catches it
    Anyone who has had a manager
  • The Black Sheep

    Where it hides
    The logo
    Who catches it
    Anyone from East Lansing
  • Shirley Jackson's "The Lottery"

    Where it hides
    The narrator's calm voice, and the Monday chore draw: a ritual nobody questions
    Who catches it
    Readers
  • Replicants

    Where it hides
    In the film, they were built for the work people didn't want. Here, they're the ones doing maintenance.
    Who catches it
    Film fans
  • Nexus-6

    Where it hides
    Nexus Diligence, the rival that got agents first
    Who catches it
    Deep-cut film fans

In the review app

Go find these yourself first. I'll wait.

  • Agents are standing by

    How to find it
    Look at any agent with nothing to do. Idle agents wear headsets.
    Why it's there
    The late-night phone bank, staffed by Codex agents. Quiet replicants mean the queue is clear.
  • The agent who won't stop pitching

    How to find it
    Click the same agent four times, fast.
    Why it's there
    Keep poking and it starts selling itself like Chip Vantage, the infomercial host.
  • BUT WAIT

    How to find it
    Click the Codex Reviewer three times.
    Why it's there
    It can hold up a BUT WAIT sign, because there's always more.
  • SET IT AND APPROVE IT!

    How to find it
    Approve three PRs in a row.
    Why it's there
    The replicants break into the infomercial's chant, the corrected version, with you doing the approving.
  • Set it and forget it

    How to find it
    Type the word forget.
    Why it's there
    The replicants start chanting the old slogan and a banner corrects them. It's the one takeaway I never want anyone to leave with. A person always approves.
  • A Voight-Kampff test on the replicants

    How to find it
    Enter the classic cheat code: โ†‘ โ†‘ โ†“ โ†“ โ† โ†’ โ† โ†’ B A
    Why it's there
    The film's test for what's real, run on the agents for once.
  • The fishbowl

    How to find it
    Click the fishbowl in the Queue room.
    Why it's there
    It held the tickets for the old Monday chore draw. Nobody draws chores anymore. Nobody threw out the bowl either.
  • WORLD'S OKAYEST SRE

    How to find it
    Hover the mug in the Codex review room.
    Why it's there
    It's Earl's, from the infomercial. He's been on call for 31 years and has earned a modest mug.
  • Word is they hired replicants

    How to find it
    Hover or focus the Nexus Diligence window across the street.
    Why it's there
    The rival's eNPS went from +12 to +48. Office gossip has a theory.

In the infomercial

Every gag is built on a true claim. The joke is the scale.

  • Chip Vantage

    Where
    The pitchman in the fleece vest.
    Why
    Every infomercial needs a host who is far too excited about lockfiles.
  • Earl

    Where
    The caption reads "31 years on call, has seen things." He wipes away a tear when the billing module finally gets upgraded.
    Why
    A nod to Roy Batty's last speech in Blade Runner, from a man holding a WORLD'S OKAYEST SRE mug.
  • Set it and APPROVE it

    Where
    The studio audience chants "set it and forget it." Chip stops them.
    Why
    The famous slogan is set up only so it can be corrected. Nothing merges without a person.
  • Agents are standing by

    Where
    A room of agents in headsets with nothing to do, idly spinning pens.
    Why
    The classic call-now line, staffed by the replicants.
  • Reviewers are standing by

    Where
    Right after the agents. It's just Leon, alone, with a coffee.
    Why
    The real bottleneck, in one shot. That's why Codex reviews first.
  • AS SEEN IN #ENG-GENERAL

    Where
    The badge on the end card.
    Why
    The only press coverage an engineer trusts.
  • Call now: @codex review

    Where
    The phone number on the end card.
    Why
    It's the one call to action that works. Comment it on a pull request.
  • Replicants not included

    Where
    The last line of fine print that scrolls far too fast to read.
    Why
    Pause it. The rest is true too, including "No fishbowls were harmed."

Sources

The research behind the note.

  • Chainguard, Engineering Reality Report, via BetaNews1,200 engineers and tech leaders. Engineers spend 16% of their week building features, and 93% call it the most rewarding part of the job. 72% say demands on their time make it hard to find room for new features. 66% of tech leaders worry about keeping engineers.
  • Faros, key takeaways from the DORA 2025 reportAI-heavy teams merge 98% more PRs while review time rises 91%.
  • DORA, 2025 reportAI adoption at 90%, with only 3% reporting high trust in it.
  • Stack Overflow Developer Survey 202546% of developers distrust the accuracy of AI tools.
  • LeadDev, 2026 survey of 600 engineering leadersMaintainability (73%) and output quality (71%) are the top concerns. Internal AI adoption is the top priority (77%).
  • Atlassian, State of Developer Experience 2025Finding information is the top source of developer friction. Half of developers lose about 10 hours a week to it.