One-offAgent Team HQ
Start with five specialists and a lead.
One job. One workflow. A team to help.
You’re in good company.
2,500+graduates
20+cohorts
15+countries
Professionals, consultants and founders. Learning by building.
Three ways to build with us.
One-off
Four weeksAgent Native Workforce
Inherit a working org. Build your own department.
Learn the system by making it yours.
AdvancedAgent Native Camp
Build the builders.
An agent workforce that builds and improves itself.
Inherit a working agent org.
Make it yours.
Four weeks. Learn the system by working inside it.
Hire your chief of staff.
Six departments.
Your own department.
Three months of The Lab.
Give Astra a job to finish.
Tell it what you want. Give it the relevant material.
Explain what done looks like.
01A clear outcome
02Useful context
03Room to work
You choose the destination.
More of the job. Less waiting.
Using a computer, across a set of test tasks.
GPT-5.6 Sol
75 minutes
GPT-6 Astra
40 minutes
Approximate time per task in OpenAI’s OSWorld 2.0 latency simulation.
For us, this feels full-stack: it can build what you see and the code underneath.
The intro started with a choice
“same thing in the same layour but 10 difernt ways?”Hunter, actual direction
Six pieces made the opening
Astra directs the tools.
Write the prompts.
ImageGen makes the visual assets.
Write the code.
Arrange the pieces. Animate the opening.
Direct the audio.
ElevenLabs makes four effects and the music.
Harness: the tools and workflow that let the workforce act.
Then the pieces started moving
Ten seconds, shown in twenty frames.
Even the spin got its own sound
Music and custom effects made with ElevenLabs.
Under the full ten seconds
A little direction changes the result.
“keep the lime in the middle but make it saras name and picture and then move me over to the left and have it in the orange”Hunter, actual direction
One clear correction.
The result changes.
⇅
Swap Hunter and Sara
What makes this work?
The opening film
Direction + governance
Hunter chooses the look.
Knowledge + memory
Brand, portraits and references.
Context engineeringAgent workforce
Workers handle parts of the job.
Skills, tools + harnesses
Image prompts, code and sound tools.
Models
Astra directs the production.
Evals + improvement
Watch it. Adjust the timing.
The agent directs the tools. Images become moving layers. Sound gets its own prompts and timing.
The Agent Workforce behind this deck
Agent Workforce: people, roles, tools and review. Solid: used here. Dotted: available roles.
What makes this work?
Direction + governance
A lesson people can follow.
Knowledge + memory
Sara’s research and our examples.
Context engineeringAgent workforce
Build, audio and review workers.
Skills, tools + harnesses
HTML, diagrams and browser checks.
Models
Astra carries the build forward.
Evals + improvement
Inspect every card. Revise.
A workforce makes the whole thing. Research, examples, media and review become one teachable lesson.
A London you can walk through
Ben Davis built a world
he could walk through.
An entire body you can pull apart
Human Atlas
by @ashebytes
Explore the layers, then watch them come apart.
Original postExplore the appApp preview: Human Atlas / BodyParts3D (CC BY 4.0)
An editor for us and our agents
Agents make the rough cut.
You make the adjustments.
Keep a useful correction as a check for the next edit.
What makes this work?
Video Studio
Direction + governance
Make one requested change.
Knowledge + memory
The project keeps its layers.
Context engineeringAgent workforce
A person and agent share the work.
Skills, tools + harnesses
Editor controls and saved edits.
Models
Astra works through the editor.
Evals + improvement
Check the picture, timing and sound.
Make “done” something you can see. Check the edit. Keep useful corrections for the next pass.
Our agent team became a game
What makes this work?
Agent Ant Farm
Direction + governance
Choose how to meet your team.
Knowledge + memory
Give each role useful context.
Context engineeringAgent workforce
Meet the workforce through roles.
Skills, tools + harnesses
Rooms and controls become a harness.
Models
Choose the model behind the work.
Evals + improvement
Check the work, beyond the game.
Choose the interface that works for you. A room, a chat or a game can be a way into the workforce.
How the agent team works.
Memory: keep useful context and reviewed lessons. The workforce carries them into the next job.
Choose the model for the job.
Build something difficult.
Give Astra the work that needs its capabilities.
Run it every week.
Try a cheaper model. Check whether the result is good enough.
Keep your context.
Your files, standards and team can stay when the model changes.
Coding test: Astra used about one-third of Sol’s tokens, but cost about 15% more per task.
Artificial Analysis Coding Agent Index, September 9, 2026. Astra at max effort versus GPT-5.6 Sol.
What stays useful
when the model changes
Direction + governance
Set the purpose and the boundaries.
Knowledge + memory
Carry useful context into the next job.
Context engineeringAgent workforce
Give people and agents clear roles.
Skills, tools + harnesses
Give the team ways to act.
Models
Choose the engine for the work.
Evals + improvement
Check what came back. Improve it.
The weekly update is a job too
This week’s notes
A few decisions.
A changed deadline.
A task list.
A draft you can use
What changed.
What’s next.
What needs an answer.
The next week starts with the same context.
What makes this work?
The weekly client update
Direction + governance
A useful draft, ready for review.
Knowledge + memory
Notes, tasks and your voice.
Context engineeringAgent workforce
Drafting and review have clear jobs.
Skills, tools + harnesses
A repeatable brief and workflow.
Models
Use a model that fits this task.
Evals + improvement
Flag dates. Check facts and length.
Next week starts with useful context. Reuse the workflow. Add fresh notes. Keep what you learned.
What’s your version of this?
One job.
Your material.
A clear finish line.
Try it with something
you actually need.
