/research · field guide 001 · updated 18 July 2026

The spec-driven development landscape

Thirteen approaches to preserving intent, directing coding agents, and checking what gets built. They share a label, but solve four materially different problems.

Disclosure: MadeForMachine builds MFM Spec, which is included below and clearly marked. We do not rank the field or award a winner. Product claims link to public documentation or repositories; boundaries and fit are our analysis.

$ one label, four jobs

“Spec-driven” now covers everything from product evidence to agent orchestration. Start with the job you need done; feature lists only become useful after that distinction.

system intent01

Capture why a system exists, what it must do, and the evidence or structure behind that intent.

MFM Spec · IntentSpec · Auto

change workflow02

Turn a proposed change into requirements, design, tasks, and enforceable implementation boundaries.

OpenSpec · Spec Kit · Kiro · SpecDD

agent execution03

Coordinate agents and people through planning, implementation, review, and delivery.

BMAD · cc-sdd · Intent · Colign · SpecForge

derived intelligence04

Recover intent and dependency context from an existing codebase instead of authoring it upstream.

ByteBell

$ comparison

“Validation” means deterministic checks on the specification—not an agent saying the prose looks plausible. “Verification” asks whether the implementation satisfies the intent. Those are separate capabilities, and few products do both.

solutioncanonical artifactspec validationpersistence / historyagent contextcode verificationportabilityprimary fit
MFM Spec(ours)typed system graphgraph + policyGit or immutable hosted revisionsprojections + MCPnot yethighdurable system intent
OpenSpeccurrent specs + change deltasschema + structureGit, archive, beta storesfiles + CLI JSONagent-ledhighbrownfield change control
SpecDDcolocated contract DSLeditor + workspace rulesGitinheritance + referencesseparate testshighsource-adjacent boundaries
GitHub Spec Kitspec → plan → tasksworkflow + extensionsGitper-feature filesimplement command + extensionsvery highgeneric spec-driven SDLC
Kirorequirements → design → tasksanalysis + hooksrepo + Kiro surfacesfeature files + steeringproperty tests + hooksmediumintegrated spec-to-code
IntentSpecevidence-backed intent fileJSON Schema + CIGitper-feature filedeclared instructionshighproduct evidence
BMAD-METHODPRD + architecture + storiesagent reviews + gatesGit artifactsrole/story contextreview + testshighfull delivery lifecycle
cc-sddrequirements → design → tasksagent cross-reviewGitbriefs + skillsTDD + reviewerhighlong autonomous execution
Intentworkspace spec + tasksworkflowapp workspace + Gitnotes + coordinatorverifier + browsermediummulti-agent supervision
Auto / NDDdomain → narrative → momentpublic details limitedhosted productbuild instructionspublic details limitedmethod: high; product: lowline-of-business behavior
Colignmemory + proposals + tasksrequired fields + workflowdatabase + Git contextMCPagent workflowhigh via MCPteam/agent alignment
SpecForgespec → epic → ticket DAGDAG + quality gateshosted lifecycleMCPevidence gatesagent: high; data: uncleargoverned execution
ByteBellcode-derived intent graphre-derived graph checksincremental IRMCPcode/change analysishigh via MCPbrownfield understanding

“High portability” means the core artifact or agent interface can move between tools. It does not imply identical behavior in every host. Publicly undocumented capabilities are marked as limited or unclear, not assumed absent.

$ field notes

What each approach actually stores, where it is strongest, and where its responsibility ends.

GitHub Spec Kit

change workflow

open source

A portable workflow for taking a feature through Specify → Plan → Tasks → Implement. Its constitution and agent commands impose process discipline while generated Markdown files stay in the repository.

artifact
A constitution plus per-feature specification, implementation plan, task list, research, and supporting contracts.
distinctive
The broadest agent integration surface in this field and an active extension ecosystem. It is a strong default when a team wants one recognizable spec-to-code ceremony across different coding agents.
boundary
The core validates workflow and artifact presence more than system-wide semantics. It is a process harness, not a typed model of component ownership or cross-feature relationships.
OpenSpec

change workflow

open source

A change-management format built around a useful split: specs describe current truth, while change folders carry proposals, designs, tasks, and delta specs. Archiving a completed change merges its deltas into the canonical specs.

artifact
Repository files under specs/ and changes/, with ADDED, MODIFIED, and REMOVED requirement deltas and scenario-based acceptance detail.
distinctive
Structural validation, machine-readable CLI output, high agent portability, and a beta store model for requirements shared across repositories. It is the closest general-purpose file-based alternative to durable spec memory.
boundary
Its center of gravity is requirements and change lifecycle, not a typed system graph. Rich component topology, revision APIs, and minimal projections are not its default abstraction.
Kiro

change workflow

commercial product

An integrated IDE, CLI, and web environment whose specs move from requirements to design to implementation tasks. Requirements can use EARS notation, and steering files supply durable project conventions.

artifact
Per-feature requirements.md, design.md, and tasks.md files, supplemented by repository steering and automation hooks.
distinctive
It connects specification to implementation correctness more tightly than most competitors. Optional property-based testing derives executable properties from requirements, while hooks can make checks blocking.
boundary
The workflow is optimized for Kiro surfaces, so the full experience is less portable than a format-first tool. Durable system modeling and independent hosted revision history are not the primary proposition.
SpecDD

change workflow

developer tooling

A source-adjacent DSL that places .sdd contracts beside code. Specs declare purpose, ownership, allowed reads and writes, dependencies, interfaces, prohibitions, scenarios, tasks, and completion conditions.

artifact
Hierarchical .sdd files: vertical rules inherit through the filesystem and explicit references establish horizontal relationships.
distinctive
Its editor integrations provide strict diagnostics and workspace validation, making behavioral and authority boundaries visible where developers work. The format is unusually good at saying what an agent may not modify.
boundary
It treats Git and the source tree as the persistence layer. It does not provide a hosted revision store, system-level projections, or a separate collaboration plane.
IntentSpec

system intent

open format

A per-feature intent format that starts with product evidence instead of an implementation plan. It records objective, evidence with anchors, outcomes, constraints, scope, edge cases, health metrics, and verification instructions.

artifact
A portable intent.md file using YAML frontmatter and Markdown, backed by JSON Schema and a GitHub Action.
distinctive
It makes the provenance of product decisions first-class. That is valuable when the important question is not only what to build, but which observed user or business signal justifies it.
boundary
The unit is a feature intent document rather than a whole-system graph. Architecture ownership, dependency integrity, and coordinated implementation are intentionally outside its core.
BMAD-METHOD

agent execution

open source

A large AI-assisted software-development method spanning analysis, product planning, architecture, epics, stories, implementation, and review. Specialist agent roles guide each phase and produce the next phase's context.

artifact
PRDs, architecture documents, epics, stories, project context, and a compact bmad-spec kernel covering why, capabilities, constraints, non-goals, and success signals.
distinctive
It offers the most comprehensive end-to-end ceremony in this set. Teams wanting repeatable roles and checkpoints across a full delivery lifecycle get much more than a spec template.
boundary
The breadth produces many artifacts and agent-mediated reviews rather than one canonical typed model. It is a delivery operating system, not primarily long-lived queryable system memory.
cc-sdd

agent execution

open source

A Kiro-inspired implementation workflow packaged as portable Agent Skills. It covers discovery, requirements, design, task decomposition, long-running implementation, review, and debugging.

artifact
Repository-based requirements, design, and task files, with boundary annotations, dependency markers, file-structure plans, and cross-spec review.
distinctive
Boundary-first design and separate implementer, reviewer, and debugger roles make it suitable for autonomous work that must survive multiple sessions. Its TDD loop supplies stronger implementation feedback than document-only approaches.
boundary
The codebase remains the eventual source of truth and most quality judgment is agent-driven. It does not try to preserve a product-wide, independently queryable intent graph.
Intent

agent execution

desktop product

A desktop workspace for coordinating multiple coding agents around a Git branch and worktree. Each workspace combines a living spec, tasks, notes, context, diffs, browser state, and activity.

artifact
A workspace-level spec and task model attached to an isolated Git worktree, with coordinator, implementor, and verifier agents.
distinctive
It makes parallel-agent execution observable to a human and includes browser-based end-to-end verification. Bring-your-own-agent support includes Codex, Claude Code, and OpenCode.
boundary
The durable unit is the Intent application workspace rather than a portable formal specification. It optimizes orchestration and supervision more than semantic validation.
commercial product + open method

A product-oriented method for describing software as domains, narratives, scenes, and moments. The hierarchy expresses user goals, outcomes, business rules, examples, and should-statements before producing build instructions.

artifact
A narrative model that reads more like structured product behavior than a conventional technical requirements document.
distinctive
The vocabulary is accessible to business stakeholders and well suited to line-of-business applications. It keeps product behavior central instead of letting architecture or task decomposition dominate the conversation.
boundary
The public method is portable, but the complete workflow is a hosted product. Public documentation does not establish the same deterministic validation or open exit surface as format-first tools.
Colign

agent execution

open source + cloud

A collaborative project-memory and proposal workspace for teams and agents. Proposals move through Draft → Design → Review → Ready with problem, scope, approach, and Given/When/Then acceptance criteria.

artifact
Shared project memory, structured proposals, tasks, acceptance criteria, comments, and a Kanban workflow with real-time collaboration.
distinctive
Its MCP surface lets agents read and write the same specifications and tasks humans review in the UI. Self-hosting and an AGPL codebase offer an uncommon open path for collaborative spec software.
boundary
It is early and collaboration-centric. Its public model does not yet show the graph integrity, immutable revisions, or implementation evidence gates of more specialized systems.
SpecForge

agent execution

closed beta

A hosted execution system that turns a structured specification into epics and an atomic ticket dependency graph. Critical paths and execution waves make sequencing explicit for coding agents.

artifact
Structured specification fields followed by epics and ticket DAGs, with planning and implementation quality gates.
distinctive
Quality gates check acceptance criteria, expected files, Git state, and test evidence before work advances. Its MCP tools expose planning and execution state directly to agents.
boundary
The product focuses on the downstream execution graph rather than durable system intent. It is proprietary and in closed beta, so portability and long-term export behavior remain less clear.
ByteBell

derived intelligence

hosted + self-hosted

A code-intelligence system that derives a verifiable intermediate representation of intent, dependencies, and business logic from existing repositories, then exposes that graph to agents through MCP.

artifact
An incrementally maintained, code-derived graph spanning repositories, with dependency navigation and blast-radius analysis.
distinctive
It addresses the brownfield reality that useful intent often exists only in code. Cross-repository queries and re-derivation after changes can give an agent context without requiring a team to author a specification first.
boundary
Its direction of travel is code → inferred intent, the inverse of an authored specification. It can explain what appears to exist, but it cannot by itself preserve product intent that the code never implemented.
MFM Spec

system intent

ours · in development

MadeForMachine's format and service for durable, agent-readable system memory. It models components, features, and evaluation notes as a typed graph that can live in local files or an immutable hosted revision history.

artifact
Typed Markdown nodes with YAML frontmatter and stable ids; hosted projects expose the same model through fine-grained MCP reads and mutations.
distinctive
Deterministic graph integrity, explicit component ownership, feature-to-component traces, minimal-context projections, optimistic concurrency, and portable export. It does not run server-side inference.
boundary
MFM Spec deliberately stops before planning and executing the change. It does not yet verify that code conforms to the graph, and the hosted service is still moving through release-candidate development.

$ shortlist by job

A pragmatic starting point, not a verdict. Most mature teams will combine layers—for example, product intent plus a change workflow plus ordinary tests.

Portable change specsOpenSpec
Broad agent ecosystemGitHub Spec Kit
Integrated spec-to-code checksKiro
Local enforceable contractsSpecDD
Evidence-backed product intentIntentSpec
Full AI delivery methodBMAD-METHOD
Long autonomous implementationcc-sdd or SpecForge
Collaborative human/agent reviewColign
Multi-agent desktop supervisionIntent
Business-readable narrativesAuto
Code-derived cross-repo meaningByteBell
Queryable system intent independent of codeMFM Spec

$ what we learned

The market is converging on Markdown artifacts and agent commands, but not on authority. Some tools treat the spec as temporary input to code; others treat it as current product truth; a smaller group treats it as durable operational memory.

The meaningful frontier is closing the loop without confusing its layers: deterministic spec integrity, economical agent context, and evidence that running code still satisfies the declared intent.

$ method

We reviewed first-party product pages, documentation, repositories, and format references available on 18 July 2026. The table compares documented product behavior, not roadmap promises. Categories, boundaries, and selection guidance are MadeForMachine editorial analysis. We will revise material inaccuracies as the field changes.

inspect MFM Spec →