Field notes on building with agents and humans
How to put agents to work without handing them the keys, how to design a pipeline that holds, and how teams ship faster while keeping a human on every merge. Written for the people doing the work.
What we write about
Supervising agents
Keeping AI fast and safe at the same time.
Pipeline craft
Designing stages and gates that earn their place.
The runtime for agents
Triggers, sandboxes, and validation in practice.
CI/CD for agents
What changes when agents open most of the pull requests.
Observability
Reading conclusions instead of noise.
Team practice
How planning, code, and operations come together.
Field notes, newest first.
Real release notes and practitioner write-ups only. No fabricated dates, no filler.
Moving to paid-only: what changes and why
JuhJuh is transitioning to a 4-tier paid-only pricing model. Here is what changed, why we made the decision, and what it means for existing users.
Building for scale: how JuhJuh handles multi-repo orchestration
Coordinating changes across multiple repositories is one of engineering's hardest problems. Here is how JuhJuh's worktree isolation model, service dependency mapping, and context flow make multi-repo development tractable.
Prompt engineering for backend developers: templates that scale
Writing effective prompts for backend development requires different patterns than frontend work. Here is how JuhJuh's template system turns prompt engineering from an individual skill into an organizational capability.
The cost of invisible AI: why token tracking matters
Your team already uses AI tools. The question is whether you know what it costs. Token-level attribution surfaces the true economics of AI-assisted development.
The 10-ticket rule: when to use Auto-Play (and when not to)
Auto-Play transforms how teams process large batches of work. But it is not always the right choice. Here is a practical framework for deciding when batch execution fits and when manual control serves better.
Why orchestration matters more than generation
Most AI coding tools optimize for the individual developer. That is the wrong unit of value. The real bottleneck is coordinating the pipeline from requirements to production across a team.
Five safety layers: how JuhJuh keeps AI out of production
Every line of AI-generated code passes through five independent review gates before reaching your production environment. Here is how the safety model works.
From idea to PR in one pipeline: the JuhJuh approach
JuhJuh connects idea capture, AI analysis, prompt engineering, code execution, diff review, and PR creation in a single orchestrated workflow. No context-switching. No manual handoffs.
The knowledge hierarchy: how your AI gets smarter over time
Four levels of agent instructions cascade from organization to project to team to developer. Every execution compounds institutional knowledge. Here is how the hierarchy works.
Auto-Play: execute an entire Epic overnight
Queue child tickets for sequential execution. Each ticket inherits context from completed siblings. Wake up to a board full of PRs ready for review.
Have a sharp take on building with agents and humans?
We publish practitioners.
