The Teammate Onboarding Model for AI Agents
Adopt a coding agent the way you'd onboard a new intern — pair first, then delegate.
- Difficulty
- Easy
- Time to result
- ~weeks to results
- Steps
- 4
- Confidence
- 90%
Rather than treating an AI coding agent as a magic delegation box, treat it like a new human teammate you onboard over time. You start by pairing interactively, learn how to prompt it, feed it context and credentials as you go, and only then graduate to handing it long autonomous work. The relationship, not the raw model, is what unlocks value.
Origin
Alexander Embiricos, Codex Product Lead at OpenAI, developed this as the core mental model for how Codex is designed and adopted, drawing on his prior work on a pair-programming startup before OpenAI.
Core principles
- 01A smart agent with no context behaves like a brilliant intern who won't read Slack or check monitoring unless asked
- 02Trust is built incrementally through side-by-side work, not granted up front
- 03Configuration (passwords, tools, guidelines) accumulates as a byproduct of pairing, not as an upfront setup chore
- 04Autonomy is the destination, pairing is the on-ramp
How to run it
- 1
Pair before you delegate
Start by working interactively with the agent inside your IDE or terminal — ask questions about the code, have it run tests and execute code with you watching, so you learn how it responds.
Pro tip Local/interactive use inside a sandbox lets the agent run commands against real dependencies and ask you when something is blocked, creating a tight feedback loop.
- 2
Learn its prompting quirks with starter tasks
Give it a few small, bounded tasks first to discover which prompts work and which don't for this specific agent, the way you'd learn to communicate with a new hire in the first couple of meetings.
Pro tip Codex prompts a little differently from other models because it's a reasoning agent — expect to recalibrate.
- 3
Feed it context and access as you work
When the agent hits a wall — a missing password, a service it can't reach, an unfamiliar convention — supply that context in the moment, exactly as you'd hand a teammate a credential or explain a guideline.
Pro tip A fresh agent is like a teammate given a brand-new computer from the store; it can't do its job until you equip it.
- 4
Delegate longer, more autonomous work
Once context has accumulated, hand off multi-step tasks — write a plan together, then let it execute for long stretches, iterating its way to a result even if the first try isn't perfect.
Pro tip Collaborate on a plan.md markdown file with verifiable steps first; agents with a verifiable plan run productively for much longer.
Watch out Don't jump straight to hands-off delegation on a fresh setup — the async-only relationship is where you want to end up, but it's hard to adopt cold.
In the wild
OpenAI's first Codex product lived in the cloud and you delegated to it asynchronously with no interactive pairing. Embiricos likens it to hiring a teammate you're never allowed to get on a call with — it works eventually, but is hard to start with.
→ OpenAI pivoted to lead with the intuitive IDE/CLI pairing experience, and Codex grew explosively (20x since August) once users could on-ramp interactively.
Common mistakes
Expecting delegation-quality output on day one
An agent with no accumulated context is a fresh hire with a store-bought laptop — it lacks the credentials, tools, and conventions to do real work, so early autonomous attempts disappoint and users conclude the tool is weak.
Skipping the pairing phase because async is the 'future'
The fully-delegated async workflow is the eventual goal, but adopting it cold is hard; teams that start there churn before they build the trust and configuration that make delegation work.
Is it for you?
Best for
Engineers and technical PMs adopting a coding agent for real production work who want durable productivity, not a one-off demo
Not ideal for
One-off throwaway tasks where you'll never build a relationship with the agent, or non-technical users with no code context to share
From the transcript
“it's a bit like this like really smart intern that like refuses to read Slack, and like doesn't check DataDog or like Sentry unless you…”
“if you hire a teammate and you ask them to do work, but they you just give them like a fresh computer from the store,…”
“as you work with them side by side, you could be like, oh, you don't have a password for this service we use. Here's the…”
From the episode
Why humans are AI’s biggest bottleneck (and what’s coming in 2026)
Alexander Embiricos (OpenAI Codex Product Lead)