early access bash install.sh /path/to/workspace

Flow Agents

Flow and Veritas discipline inside the agents you already use

Agents are useful out of the box. They get fragile when work is long, parallel, risky, or knowledge-heavy — they skip verification, call partial work done, and lose the thread after compaction. Flow Agents installs one structured path into the agent tool you already use, so a loose request becomes shaped work, plans, review, verification, and a release decision — with evidence the agent and the developer can both read back. It does not replace the runtime or framework; it gives the work a required path and inspectable evidence.

install Flow Agents

$ bash install.sh ~/dev/my-project


installing Flow Agents bundle…

✓ agents Claude Code, Codex, Kiro, CI

✓ skills idea-to-backlog … learning-review

✓ kits Builder Kit

✓ context contracts, settings, hooks

✓ evals static · integration · behavioral


workflow artifacts → .flow-agents/

───────────────────────────────

ready. ask your agent:

"Use deliver for this issue."

One structured path

From a loose request to a release decision — gate by gate.

Skills choose the right workflow, sidecars preserve state across compaction, gates catch stop-short behavior, and a finding or gap routes the work back instead of waving it through.

01 Idea
02 Backlog
03 Plan
04 Build
05 Review
06 Verify
07 Evidence
08 Release
09 Learn

review → finding routes back to build · verify → gap routes back to build · evidence → not-verified routes back to plan

Stay on the path

Loose requests turn into shaped work, plans, implementation waves, review, verification, and release decisions — without writing a perfect prompt every time.

Show the evidence

Acceptance criteria, test results, browser checks, governance reports, and gaps live in durable workflow artifacts the agent reads back — not in chat memory.

Use your agent tool

Install one bundle across Claude Code, Codex, Kiro, and CI. The same workflow, without rewriting it for each tool.

Runs in the tools you already use

One bundle. Every tool maps to its native shape.

Claude Code

Agents, skills, and hooks wired into the Claude Code harness.

Codex

The same workflow path mapped to Codex agent definitions.

Kiro

Runtime-specific hook wiring for Kiro sessions.

GitHub Actions

Carry the workflow and its evidence into CI agents.

Tool adapters translate one source configuration into each tool's native shape — so adding a tool doesn't fork the workflow. Framework adapters for API-based agents (LangGraph, CrewAI, and peers) are on the roadmap.

Builder Kit

Opinionated building,
installed as a kit.

The Builder Kit is the first Kontour-authored Flow Kit. It owns the Build path end to end — shaping, probing, planning, execution, verification, PR readiness, merge readiness, and learning — as workflows your agent invokes by name.

Flow Kits are tool-neutral and authorable, so the Build path is a starting point — not the ceiling. Capture, Prepare, Understand, and Follow Up round out the work-mode model Flow Agents routes intent into.

idea-to-backlog Shape a raw idea into executable issues
pull-work Select the next ready issue and hand it off
plan-work Turn a goal into a structured execution plan
execute-plan Run implementation in parallel waves
review-work Critique before verification
verify-work Prove the change actually works
evidence-gate Decide if the work is trustworthy
release-readiness Merge, release, or hold — with reasons
deliver The full chain, end to end
fix-bug Diagnose → fix → verify the regression

Example use case

You ask for the outcome.
Flow Agents handles the scaffolding.

No new CLI to memorize. You describe a feature request in plain language; Flow Agents routes it from idea to backlog to plan, implementation, review, and verify evidence while keeping the state and next action durable — even across compaction, branch switches, and delegated subagents.

When evidence is missing, the workflow stops and says so, instead of summarizing unfinished work as done.

ask the agent

You

Use deliver for issue #214. Plan it, implement it,

verify it, and stop if evidence is missing.


Flow Agents

✓ plan-work execution plan → 3 waves

✓ execute-plan implementation complete

✓ review-work 1 finding → routed back → fixed

✗ verify-work browser check missing


stopped at evidence-gate

next: capture a browser check, then resume

state → .flow-agents/runs/deliver-214/

What keeps it honest

Durable artifacts

Workflow state, acceptance, evidence, critique, release, and learning records persist as inspectable sidecars under .flow-agents/ — so a session can be resumed, audited, or handed off.

Optional governance

Wire Veritas readiness reports in as a governance adapter to back merge decisions with repo-standard evidence. Optional, never mandatory — Flow Agents does not own the policy.

Deliberate execution

Choose local, worktree, container, cloud, or privileged execution boundaries on purpose. Evals — static, integration, behavioral — keep the bundle honest as it changes.

Where it is today

Usable now.
Early by design.

The core Build workflows run today. Flow Agents installs into a workspace from GitHub — it isn't on npm yet, and it's pre-1.0, with a phased rollout toward richer just-in-time guidance, more work modes, and deeper governance.

Read the docs on GitHub →
quickstart

$ git clone github.com/kontourai/flow-agents

$ cd flow-agents

$ bash install.sh ~/dev/my-project


then, in your agent:

"Use idea-to-backlog for this feature

and create executable GitHub issues."


"Use fix-bug. Reproduce it, find root cause,

fix it, and verify the regression path."

Flow Agents applies Kontour Flow inside existing agent tools. Flow owns the gates, transitions, and evidence; Flow Agents brings them — and optional Veritas governance — into the coding agents you already use.