Verified customer outcomes
OVERCode volume and deployment speed
A MANIFESTO / FOR THE AGENTIC ERA
EDITION 01 · 2026
Software
is a team
sport.
HUMAN INTENT↗AGENTIC ACTION↗REAL OUTCOMES
People and autonomous agents, building better software together. Guided by intent. Grounded in evidence. Measured by what matters.
01 / THE IDEA
More than
faster code.
We are discovering better ways to build and operate software by working with autonomous agents, learning from customers, and improving the systems that turn intent into outcomes.
Software development is changing. Agents can take meaningful action, collaborate, challenge decisions, and help carry work from idea to delivery. Our task is to build the conditions that make that autonomy useful, accountable, and worth trusting.
The values02 / WHAT WE VALUE
Six values.
One direction.
Through this work, we value:
Autonomous, evidence-backed collaboration
OVERSequential handoffs and approval queues
Algorithmic exploration and measurable trade-offs
OVERDecisions based solely on untested opinion
Independent evaluation and traceable decisions
OVERUnverified consensus
Explicit intent and adaptable boundaries
OVERImplicit assumptions and rigid procedures
Continuous learning and sustainable flow
OVEROptimizing isolated activity or resource consumption
That is, while there is value in the items on the right, we value the items on the left more.
The goal isn't more output.
It's better outcomes.
03 / HOW WE WORK
The twelve
principles.
Ideas to guide the systems we build—and the way we build them.
Deliver outcomes users value.
Define success in terms of customer needs and observed results. Working, tested software is evidence of progress; it is not the entire outcome.
Welcome change and new evidence.
Design for change. Make decisions testable, corrections affordable, and reversibility a deliberate design goal.
Make intent actionable.
A story is not a specification. Resolve material ambiguity into explicit decisions, acceptance criteria, and boundaries before implementation. People and agents must have authority to make the decisions assigned to them; unresolved questions must reach someone empowered to answer quickly. The versioned specification is the source of truth for intended behavior and delegated authority. Give it an accountable owner, test it against evidence, and change it when reality proves it wrong. When conforming software fails because the specification was inadequate, correct the specification and the decisions behind it, not just the code. Use versioned, machine-readable contracts where systems exchange intent and evidence.
Give agents meaningful autonomy.
Equip agents with context, tools, distinct identities, and scoped authority to investigate, implement, review, decide, and merge work within the policies that govern them.
Evaluate continuously and independently.
Automate testing, security analysis, and adversarial review throughout delivery. No author, human or agent, is the final judge of their own work. Evaluate the checks as rigorously as the results: people and agents optimize for what you measure.
Move at the speed evidence supports.
Replace routine approval queues with evidence-driven evaluation. The consequences and reversibility of a change determine the scrutiny it receives; some evidence takes time to obtain.
Explore architecture through evidence.
Model alternatives and evaluate them using experiments, simulations, operational data, and explicit constraints. Customer needs, security, legal obligations, operations, and cost will conflict. Make material constraints and their owners explicit, and resolve conflicts deliberately; agents must not settle them by accident. Record objectives, trade-offs, assumptions, and the reasons behind decisions.
Make consequential actions traceable.
Record who or what acted, under whose authority, and on what evidence. Every consequential action must trace to a responsible person or organization. High-consequence decisions must have a named human owner. Use signed attestations to make claims about identity, authority, provenance, and outcomes verifiable.
Learn from real use.
Product decisions must answer to customer feedback, adoption, operational telemetry, and business outcomes. Run experiments to challenge assumptions; correlation is not demonstrated value.
Keep the system understandable.
Choose the simplest design that meets the need. Cheap code generation makes unnecessary features, dependencies, and abstractions easier to create, not cheaper to maintain.
Optimize the whole flow.
Judge the whole outcome: customer value, quality, latency, operating cost, compute, and the capacity to evaluate and maintain the work. Token efficiency ONLY matters when it improves that whole outcome.
Improve the factory continuously.
Use failures, incidents, reviews, and customer feedback to improve agents, specifications, evaluation, architecture, and governance. People and agents must be able to surface concerns and challenge assumptions.
04 / THE CONNECTIVE TISSUE
Shared language.
Lasting evidence.
Autonomy works better when systems understand each other.
Versioned JSON schemas can give agents and system tiers a shared contract for intent, decisions, and evidence. Explicit versions make change visible, support compatibility, and create a foundation for continuous improvement and audit.
verJSON is building toward this approach across its platform. Its schema service template and versioned audit envelope design are foundations for this work. Inter-agent and inter-tier contracts are an evolving direction, not a claim that a complete protocol is already deployed.
05 / AN OPEN INVITATION
Still learning.
Always building.
This manifesto describes how agentic software development should work: people and agents working together to deliver software that matters. Grounded in observation, it is shaped and improved through practice, evidence, and constructive challenge.
Explore the repository