Describe the intent.
Declare the needed planning, delegated work, isolation or evidence behavior independently of vendor names.
Architecture
Your delivery system.
Not your provider's workflow.
AgentFlow separates the rules for good delivery from the tools that carry out the work. Extend one layer without rebuilding the others.
Four boundaries. One coherent flow.
01 / Your project
Business-analysis methods, engineering conventions, review expectations and workflow profiles.
Configuration + method selection + extension packs02 / AgentFlow core
Roles, contracts, criteria, evidence verification and advancement gates.
Portable policy, independent of a harness03 / Provider ports
Discover capabilities and bind an execution intent to the selected harness.
Planning + delegated work + isolation, where supported04 / Source ports
Keep durable records in the workflow source. GitHub is the shipped source integration; Cockpit is an optional view.
Read and mutation boundaries + receipts + provenanceAgentFlow is not a replacement for your coding assistant, repository or technology stack. It defines how their work becomes an accountable delivery.
Read the technical architecture ↗Extensible by design
Change the method.
Keep the contract.
A business analyst can use interviews or event mapping. An engineering team can choose its planning method. The responsibility, evidence and acceptance boundaries stay explicit.
Roles define what is owned.
Canonical identities such as agentflow:analyst and agentflow:developer define responsibilities, scope, inputs, outputs and behavior. Nine lifecycle roles share one taxonomy.
Methods define how the work is approached.
Select or contribute approaches without rewriting role ownership. Team-specific analysis, design and planning practices can vary while their output contracts remain reviewable.
Skills support bounded work.
Orchestrator, collaborator, scanner, designer, migrator and auditor skills support coordination, discovery, policy design, adoption and review. Skills are not extra lifecycle roles.
Profiles and extensions tailor the product.
Choose minimal, standard, GitHub or Cockpit composition. Tune complexity thresholds and council seats. Add policy through extension packs; opt-in pack validators are trusted code, not a sandbox.
Versioned ports connect new tools.
Provider and source contracts expose narrow capabilities and evidence. Add a provider without injecting vendor-specific orchestration into the core, or add a source without changing what acceptance means.
Multi-harness, without lowest-common-denominator behavior
Use more when available.
Assume less by default.
A harness is the environment that runs an AI agent. AgentFlow asks what that environment can do now, rather than assuming every tool offers the same features.
Claude CodeCodexAgyPiGrokManual
Inspect and bind.
The selected provider reports support, fidelity, limits and provenance. The project chooses its executor and model explicitly.
Record the outcome.
Use an allowed sequential or manual fallback, or block when required support is missing. Never silently substitute another executor.
AI Foundry Desk is an optional integration, not an embedded orchestrator. Its shipped integration exposes inventory, project adapters and evidence. It does not currently advertise execution, workspace or lifecycle support. An xAI API binding requires explicit configuration.
Provider support varies by installed version and configuration. Capability discovery is not an end-to-end certification of every native feature.
Build a provider ↗What does not become optional
Adaptability needs
stable boundaries.
Customization changes the method, not the right to bypass acceptance. Keep a single accountable writer, explicit action authority and independent review where required.
No evidence, no advancement.
Stale decisions, missing required criteria, conditional acceptance and open rework block progress.
No hidden authority.
Content integrity is not identity. Source and harness authorization controls remain responsible for who can approve and act.
No automatic overwrite.
Adoption previews exact changes, checks the current token and writes transactionally. Project-owned conflicts fail closed.
Start small. Keep control.
One repository.
One clear next step.
Architecture under implementation
Shared services. Explicit evidence.
The delivery update adds application services between the portable policy core and its execution and source adapters. The CLI and Cockpit consume the same state.
Verification and recovery
Typed observations bind checks to candidates. Recovery rereads source state and transfers writer ownership only after its preconditions hold.
Environment and adoption
Separate inspection from authorized probes. Project-contained transaction storage keeps rollback evidence outside managed files.
Operations and budgets
Keep release, deployment and rollback observations distinct. Report usage fidelity and enforce only the budget controls a provider can support.
Development preview · release acceptance pending · AgentFlow requires a supported AI client and any necessary subscription or API access.