Know why work is ready.
Acceptance connects to agreed criteria and current evidence, not just a confident summary in a chat.
A reusable delivery system for AI-assisted teams
AgentFlow SDLC 2.0
Move faster. Know who owns the outcome.
Turn AI-assisted work into clear responsibilities, reviewable evidence and accountable decisions. Keep your tools. Make the process yours.
Open source · Runs locally · Works with your AI tools
AgentFlow provides the delivery process. You provide a supported AI client and any required subscription or API access.
One change. A clear path to acceptance.Interactive illustration, not a live agent run
01 / PURPOSEFrameOutcome & scope
02 / CONTRACTHand overCriteria & owner
03 / DELIVERYBuildWork & evidence
04 / VERIFICATIONCheckCurrent evidence
05 / DECISIONAcceptEvidence complete
Sender checks the delivery against the handover
Ready for sender acceptance
Required evidence is current. The next role can return its delivery for acceptance.
Routine change: deterministic checks first. No council required.
Built with AgentFlow
See a real product.
Audit how it reached production.
AwesomeAwesomeness was carried across separate delivery episodes with explicit roles, deterministic gates, fresh-context recovery and separately checked release and hosted behavior. The product is live; its complete delivery story remains with the product.
- 3delivery episodes
- 2fresh-context recoveries
- Exactrelease and hosted verification
For leaders and the teams they trust
AI changes the pace.
Accountability stays.
More generated code is not the same as a dependable outcome. AgentFlow makes it clear what was agreed, what was delivered and why work is ready to move forward.
Keep the process when tools change.
Roles, decisions and review boundaries belong to your delivery system, not to a particular AI vendor.
Spend judgment where it matters.
Use code for repeatable checks. Bring in specialist advice and human authority when complexity or risk calls for them.
What makes 2.0 different
A process you can adapt.
Not a platform you inherit.
Five connected capabilities make the delivery model reusable across teams, methods and AI environments.
Architecture
Separate the process from the machinery.
Change your execution tools without rewriting the rules that govern delivery.
HOW IT WORKS
A portable core owns roles, contracts and gates. Providers connect execution tools; source adapters connect durable records.
Explore the architectureExtensibility
Your methods. Shared accountability.
Business analysis and engineering can work differently without forking the product.
HOW IT WORKS
Versioned roles define responsibilities. Methods, skills, profiles and extension packs tailor how each responsibility is fulfilled.
See what you can customizeDeterminism
Check facts before asking for judgment.
Don't spend a model call deciding whether a required artifact exists or whether evidence has changed.
HOW IT WORKS
Code verifies required criteria, record integrity and current content. Missing evidence, stale approval and unresolved rework block advancement.
Understand the checksMulti-harness
Use the capabilities you actually have.
Benefit from planning, subagents and isolation when available, without making one AI environment mandatory.
HOW IT WORKS
Providers inspect support and bind an explicit execution target. A permitted fallback is recorded; a missing required capability blocks the plan.
Understand harness supportCollaboration
More expertise. One accountable owner.
A handover starts a working agreement. It doesn't end the sender's responsibility.
HOW IT WORKS
The receiving role returns evidence. The sender accepts conformance or requests rework. Targeted councils advise on complex decisions.
Follow the acceptance loopAn operating model, not another coding agent
One agent can wear
several hats.
A role is a responsibility, not necessarily another AI agent. Start with one executor. Add independent perspectives only when they improve a decision.
Meet the nine lifecycle roles →Your team sets the boundaries.
Define the delivery method, required evidence and risk policy. High-assurance decisions retain a human gate.
The harness does the execution.
Claude Code, Codex, Agy, Pi, Grok or an explicit manual path. Capability availability is inspected, not assumed.
AgentFlow connects the decisions.
Scope, handover, delivery, verification and acceptance remain inspectable across roles and sessions.
From 1.0 to 2.0
From a shared workflow
to a reusable system.
1.0 established the foundations: roles, evidence, validation and human review. 2.0 turns those foundations into explicit, extensible contracts.
| Capability | 1.0 / The foundation | 2.0 / The reusable product |
|---|---|---|
| Responsibilities | Named role-based phases and guidance. | Canonical prefixed roles, defined parameters and independently selectable methods. |
| Handover | Durable context and role-pass evidence. | Bilateral acceptance contracts, delivery receipts and bounded rework. |
| AI environments | Harness-neutral workflow and native adapters. | Portable intents, inspected capabilities and explicit provider binding. |
| Collaboration | Optional advisers and multi-agent evidence. | Rule-based complexity paths and structured councils with an accountable owner. |
| Adoption | Review-first installation and ownership-aware updates. | Preview-first transactions, current-plan approval, external receipts and rollback. |
A deliberate clean break. Legacy commands and v1 lockfiles are not supported by 2.0. Evaluate in a separate checkout and review existing project state before fresh adoption. The 2.0 implementation is available from source; a versioned npm 2.0 package has not been published.
Moving from 1.0? Start here →Start small. Keep control.
One repository.
One clear next step.
Delivery improvements in development
Know what passed. Recover what stopped.
The next 2.0 delivery update connects observed test results, recoverable runs and a clearer operator view. These additions are being validated before release.
Evidence tied to the work
Checks identify the exact candidate, invocation and expected behavior. A success message alone cannot satisfy acceptance.
Recovery with context
Resume from recorded state after checking the workspace, prior writer and unfinished external actions.
A visible next step
See missing evidence, open findings and the next eligible action through the CLI and optional Cockpit.
Development preview · release acceptance pending · AgentFlow requires a supported AI client and any necessary subscription or API access.