How it works

Up to 7 phases. Scaled to the work.

The platform classifies intent complexity and right-sizes the pipeline — from a single-agent one-shot to the full 7-phase SDLC. Simple tasks get 1 phase. Moderate changes get 3-5. Complex multi-repo features get all 7 with parallel reviewers, planning, and verification gates.

The pipeline

From intent to pull request.

  1. 01
    Intake

    One door in.

    Every feature request, bug, migration, or refactor begins the same way — as a plain-language intent from Slack, Jira, the dashboard, or the API. The platform enriches it with AI, classifies scope, and produces a structured work item with confidence scoring and budget checks.

  2. 02
    Discovery

    Understand before building.

    AI-driven codebase research with per-repo freshness tracking. Impact analysis maps which repositories and files will be affected. Constraints, conventions, and prior decisions are pulled from the knowledge base and fed into downstream phases. This is the phase most AI tools skip entirely.

  3. 03
    Review

    Five parallel reviewers before code is written.

    Architecture review, security review, design review, verification strategy review, and delivery review run in parallel. Each produces a structured verdict: approve, approve with conditions, or reject. If any reviewer rejects, the pipeline halts. Code is not written until the plan has been reviewed.

  4. 04
    Planning

    Structured plan with verification gates.

    The platform constructs an execution plan from the reviewed work — what gets built, in what order, with structured checkpoints between each step. The plan is validated for conflicts and impact before execution begins.

  5. 05
    Execution

    Specialized agents. Resilient execution.

    Purpose-built agents execute each step of the plan. The runtime is model-agnostic — Claude, Llama, Mistral, or Nova, selected per stage. Automatic recovery from failures. State preserved at every boundary. Live progress streaming to the dashboard.

  6. 06
    Verification

    Pass or halt. No exceptions.

    Structured verification verdicts: pass, fail, flaky, blocked, insufficient evidence. Tests run. Contracts validate. Quality gates enforce. If a step fails, the pipeline stops, logs the failure, and escalates. Nothing proceeds until the previous step has passed.

  7. 07
    Delivery

    Review-ready pull requests. Full audit trail.

    Coordinated pull requests across every affected repository. Documentation generated. Changelogs written. The full Activity Stream is queryable — every agent decision, every state transition, every human override. The work arrives review-ready. Production-readiness is your review's decision.

The product

What a running execution looks like.

Horizontal DAG on the left. Live agent feed on the right. Every agent decision streamed in real time with cost tracking.

Swarm — "Assign team members to appointments"
Running
00:34:12$2.41
Intake
Discovery
Review
Planning
Execution
Verification
Delivery
Intake
intake-interpreter
Research
codebase-researcher
Requirements
requirements-architect
Architecture
architecture-reviewer
Security
security-reviewer
Design
design-reviewer
Plan
task-planner
Backend
backend-engineer
Frontend
frontend-engineer
Mobile
mobile-engineer
Verify
qa-verification
Deliver
delivery-handoff
Live agent feed
backend-engineer34:12

Implementing POST /api/appointments/team-members endpoint

frontend-engineer34:08

Creating TeamMemberAssignment component with form validation

backend-engineer33:41

Added AppointmentTeamMember model with foreign keys

frontend-engineer33:22

Reading existing appointment detail page patterns

task-planner32:58

Execution plan approved — 3 parallel tracks

security-reviewer31:15

Approved — no new auth surfaces, existing RBAC sufficient

architecture-reviewer31:02

Approved with condition: add index on appointment_id

design-reviewer30:44

Approved — consistent with existing team management patterns

codebase-researcher28:30

Mapped 4 affected files across 3 repos, 2 existing patterns

7 complete 2 running 3 pending
Standard tier · 12 agents · 3 repos
Engineering under the hood

Distributed-systems engineering for agent execution.

Crash recovery, distributed locking, liveness detection, immutable audit trails. These are the mechanisms that make consistent output possible at scale.

Deterministic flow

Every work item follows the same path through the system. No shortcuts, no special cases. Consistent output from consistent process.

Structured checkpoints

Between each phase, structured contracts define what must be present before the next phase begins. Validated programmatically — no soft handoffs.

Built for reliability

Automatic recovery from failures. State preserved at every phase boundary. Stalled agents detected and restarted. Nothing is lost if something goes wrong.

Immutable audit trail

Every decision recorded. Queryable by execution, phase, agent, or outcome. Independently verifiable — your auditor can confirm the chain with no Spire access required.

Human-in-the-loop

Automation that works for your engineers.

Spire handles the implementation work that should not have required your engineers in the first place — migrations, framework upgrades, test backfills, API versioning, vulnerability patches. Your team stays in control at every boundary.

Plan review

Before code is written, the execution plan can be adjusted. Phases can be added, removed, or restarted. The graph is editable.

Approval gates

Configurable checkpoints where human approval is required. The pipeline waits. State is preserved. The reviewer sees full context.

Live intervention

Pause, redirect, or abort at any time. Escalations route to the right reviewer. Human judgment is a first-class input.

Post-delivery feedback

Multi-round feedback conversation to refine output. The feedback agent has full workspace access — reads, edits, runs tests, commits, and pushes. Costs tracked per round.

Additional capabilities

Beyond the pipeline.

The pipeline is the core. These are the surfaces that make it usable at enterprise scale.

Enterprise knowledge

Upload runbooks, design docs, policies, and wiki pages. Agents retrieve this context at execution time — reasoning over your documentation and your code. Designed for legacy codebases where institutional knowledge matters as much as the source.

Native Bedrock runtime

Runs natively on AWS Bedrock. No external dependencies. Model-agnostic — Claude, Llama, Mistral, or Nova at each pipeline stage. Fast, reliable, and everything stays in your VPC.

Enterprise SSO

Admin-configurable SAML 2.0 and OIDC from the settings dashboard — Entra, Google Workspace, Okta, or any provider. Under 5 minutes to configure. Standards-based security. Break-glass local admin always preserved.

Codebase Copilot

Project-integrated conversational AI. Searches your repos in real time, retrieves enterprise docs, and maintains context across long conversations. Propose work, review cost/time estimates, confirm, and watch execution — all in one continuous thread. Produces artifacts: PDFs, mockups, documentation, data exports. The conversation is the whole experience.

Integration ecosystem

Slack /spire commands, Jira sprint imports, MCP server connectors for databases, APIs, and external tools. Agents use your team's tools directly.

Vulnerability management

Import security findings, associate Jira tickets, and run structured fixes through the pipeline. Evaluate, dismiss, or fix — each action flows through the same path.

See it run on your code.

Bring a feature request from your backlog. We run it through the pipeline on your codebase. You see the pull requests at the end.