A tool-enforced multi-agent pipeline architecture for Kilo Code. Decompose complex software engineering tasks into a structured pipeline of specialized subagents — triage, context-gather, design, implement, integrate, review, remediate, and deliver — with every stage enforced by the Kilo Task() tool.

The VS Code extension is built on the Kilo CLI and reads the same on-disk config files. Pick either platform — or both — and the same agents work.

1. Open Settings → About → Import Settings
2. Select vscode/agents.json
3. Save → ceo is now your default agent
4. Start a session: @ceo <your task>

cd /path/to/your/project
cp -r path/to/kilo-cli/.kilo ./
kilo
@ceo <your task>

flowchart TB
    U["User Request"] --> CEO

    CEO["CEO (orchestrator)

    Routes to the right stage by
    delegating to specialized
    subagents via the Task() tool"]

    CEO --> S0["0. Requirement Triage

    Classify task risk and depth.
    Determines pipeline level:
    BOUNDED or COMPLEX."]

    S0 --> S1["1. Context Engineering

    Gather and synthesize context
    from repo structure, code
    history, and external sources."]

    S1 --> S2["2. Solutions Architecture

    Design a technical plan from
    the Context Brief. Produce a
    Design Document for review."]

    S2 --> S3["3. Implementation

    Execute the design in atomic,
    verifiable slices. Each slice
    includes verification evidence."]

    S3 --> S4["4. Integration

    Connect slices together.
    Check cross-file consistency,
    imports, and interfaces."]

    S4 --> S5["5. Independent Review

    Multi-track review gates:
    QA / Fidelity / Security /
    Performance."]

    S5 -- "findings" --> S6["6. Remediation

    Fix review findings with
    minimal blast radius.
    Re-run verification per fix."]

    S6 -- "gates pass" --> S7["7. Delivery

    Verify acceptance criteria.
    Clean up temporary artifacts.
    Confirm delivery."]

    S6 -. "re-enter review" .-> S5
    S7 --> DONE["✅ Delivered Result"]

    style U fill:#1e293b,stroke:#0ea5e9,stroke-width:2,color:#fff
    style CEO fill:#eef2ff,stroke:#6366f1,stroke-width:2,color:#1e293b
    style S0 fill:#e0f2fe,stroke:#3b82f6,stroke-width:2,color:#1e40af
    style S1 fill:#e0f2fe,stroke:#06b6d4,stroke-width:2,color:#155e75
    style S2 fill:#ede9fe,stroke:#8b5cf6,stroke-width:2,color:#4c1d95
    style S3 fill:#dcfce7,stroke:#22c55e,stroke-width:2,color:#166534
    style S4 fill:#fef9c3,stroke:#eab308,stroke-width:2,color:#854d0e
    style S5 fill:#ffedd5,stroke:#f97316,stroke-width:2,color:#9a3412
    style S6 fill:#fecaca,stroke:#dc2626,stroke-width:2,color:#7f1d1d
    style S7 fill:#d1fae5,stroke:#10b981,stroke-width:2,color:#065f46
    style DONE fill:#064e3b,stroke:#10b981,stroke-width:3,color:#a7f3d0

Each stage is enforced by the Task() tool — skipping a stage produces an error. Review findings are resolved through structured remediation loops (remediation → re-enter review) before delivery.

kilocode-agents/
├── README.md                       ← You are here
├── LICENSE / LICENSE-CODE / NOTICE / AUTHORS
│
├── vscode/                         ← VS Code extension users
│   ├── agents.json                 ← Bundle for Settings → Import
│   ├── agent-imports/              ← Per-agent files
│   └── README.md
│
├── kilo-cli/                       ← Kilo CLI users
│   ├── .kilo/
│   │   ├── agent/                  ← Canonical agent definitions (.md + YAML)
│   │   ├── command/                ← Optional slash commands
│   │   ├── skills/                 ← Optional pipeline skills
│   │   └── kilo.jsonc
│   ├── scripts/
│   │   ├── README.md               ← Platform-specific script guide
│   │   ├── posix/                  ← Linux, macOS, WSL, Termux
│   │   │   ├── setup.sh
│   │   │   └── generate-json.sh
│   │   └── windows/                ← Windows (PowerShell)
│   │       └── setup.ps1
│   └── README.md
│
└── shared/                         ← Cross-platform reference
    ├── context/
    │   ├── context-brief-template.md
    │   └── design-doc-template.md
    ├── old-prompt.md
    └── new-prompt.md

The canonical source of truth is the .md files in kilo-cli/.kilo/agent/. JSON artifacts in vscode/ are derived from these files. See kilo-cli/README.md for the rationale.

1. Tool-enforced pipeline. Stages are enforced by Kilo's Task() tool, not by prompt text. Skipping a stage produces a tool error.

2. Context is first-class. Every task starts with context gathering. The pipeline treats context engineering as a distinct phase with dedicated tooling — not an ad-hoc step buried in a prompt.

3. Separation of concern. Each pipeline stage has its own agent with a focused role: Context Engineer gathers, Architect designs, Implementer builds, Reviewer verifies, Remediator fixes.

4. Transparent review. Every significant decision passes through visible review with explicit approval gates. Multiple review lanes (QA, fidelity, security, performance) for different risk dimensions.

5. Structured remediation. Review findings feed a dedicated remediation loop. The remediator fixes issues, re-runs verification, and re-enters the review gate until all checks pass.

• Kilo CLI v7.3.12 or newer
• A capable model: Claude Opus, GPT-5.4, DeepSeek V4, or similar
• For CLI users: npm install -g @kilocode/cli or binary from GitHub releases
• For VS Code users: Kilo Code extension (pre-release) from the marketplace

Why markdown files instead of JSON? The Kilo CLI auto-discovers agent definitions from .kilo/agent/*.md files. Markdown is the canonical format; JSON (vscode/agents.json) is a derived artifact for VS Code Settings import.

Can I use both CLI and VS Code? Yes. The VS Code extension is built on the Kilo CLI backend. Both read the same .kilo/agent/*.md files from the same on-disk paths. The .kilo/ directory is portable.

How do I customise an agent? Override any agent by creating a .kilo/agents/<name>.md file in your project or by adding it to the agent key in kilo.jsonc. Properties are merged — you only need to specify the fields you want to change.

How do I disable an agent? Set disable: true in the agent's frontmatter or in the JSON config:

{ "agent": { "some-agent": { "disable": true } } }

What model should I use? The ceo agent benefits from the strongest model available. Sub-pipeline agents can use a cheaper model — set subagent_model in kilo.jsonc.

See CONTRIBUTING.md.

See CODE_OF_CONDUCT.md.

Prompts and agent architecture: CC BY 4.0. Code and configuration: Apache-2.0. See LICENSE and LICENSE-CODE.