Campo private beta
Request early access

CAMPO · Autonomous build agent

Set the goal. Let it ship.

Campo is an AI team that builds your application over days, fixes its own mistakes along the way, and proves the result works before saying done. Not a prototype — the actual app.

Request early access

Builds the actual thing.

· Where Campo fits

Three ways AI writes code.

Ad-hoc agent Workflow pipeline Campo
Run style Minutes, you in the loop Fixed DAG, no reasoning between nodes Days, autonomous
Input Single edit or query Trigger event into a pre-built graph Product goal / spec / concept
Output Code suggestion or one task done Outputs from each node Working, bench-passed application
Best for Quick edits, exploration, learning the codebase by hand Predictable repeating tasks with known structure Multi-step product goals you want shipped without driving
Less suited Days-long autonomous progress Task structure isn't known up front You want to drive each step

What Campo handles

  • Production-grade apps — web, non-web, plus the GUI/UX that goes with them.
  • Refactors and features in existing code.
  • Concept → working app.

What Campo decides on your behalf

  • Architecture for the build — you bring the goal, Campo plots the route.
  • Code review → replaced by bench-checks + a human-judge moment at QA.

Where you stay in the loop

  • Product strategy and roadmap.
  • Marketing and other non-engineering work.
  • Step-by-step driving — Campo shows intermediate steps but isn't a per-move assistant.

§ 01 Hierarchy

Goals descend.

Strategic → Tactical → Operational → Executive.

Each layer knows only what it needs to. No agent drags the whole project through every decision.

findings technology 3 architecture 2 entry point 1 sandbox probe python 3.13 ✓ pytest ✓ node 22 ✗ → evidence written

§ 02 Discovery

Campo reads the
project first.

Before a single action — walk the files, probe the sandbox, write everything to evidence.

Δ prediction observation

§ 03 Prediction

Predict, then verify.

Each agent predicts evidence before acting. Surprises rise.

working 7 slots summary rolling evidence ledger persistent per workspace

§ 04 Memory

Memory
in layers.

Working · summary · evidence · persistent. Per workspace.

Facts, not guesses. Each layer accumulates evidence scoped to its work; neighboring layers are one peek away.

goal patch escalate prediction error 0.74 severity > 0.50 threshold memory failure + recovery, kept

§ 05 Anomaly

Nothing fails silently.

An action breaks its prediction. Campo scores the error, escalates it up the chain, grows a working patch — then keeps the whole episode in memory.

Stops only on success. An open anomaly blocks done until reality catches up to the plan.

L1 GET cli file qa L2 GET cli file qa L3 GET cli file qa fix → rerun L4 GET cli file qa run verified

§ 06 Benchmarking

Done means proven.

Every requirement becomes a graded check — HTTP, CLI, files, browser runs. Campo climbs the ladder, fixes what fails, reruns until the level passes.

Not “I think it works” — the bench says it does.

§ 07 Workspaces

Each project,
its own brain.

  • 01

    Sandbox isolation.

    Agents run inside gVisor — one project's files, nothing else.

  • 02

    Per-project memory.

    What one project learns never bleeds into another.

  • 03

    Real-time stream.

    Watch every agent decision the moment it's made, and steer mid-run.

§ 08 Request

Request early access.

Private beta — invite by approval. Tell us what you’d want Campo to orchestrate first; we’ll reach out within a few days.

We’ll only use this to reach out about the beta. No list resale, no marketing.