Snowflake-Labs/cocoplus

What is CocoPlus?

CocoPlus is an Agentic Operating System built on top of Snowflake’s Coco platform. It adds a full‑stack, multi‑agent workflow engine for data‑engineering projects, guiding a project from the initial spec all the way through to production shipment. The system is made entirely from Coco‑native building blocks – Skills, Agents, Hooks, and AGENTS.md – so everything runs inside the Snowflake environment.

---

Core ideas

Concept	What it means in plain language
Agentic OS	Think of a tiny “operating system” that coordinates many specialised bots (agents) to do parts of a data‑engineering job, rather than a single script doing everything.
Structured lifecycle	Every project follows the familiar phases: spec → plan → build → test → review → ship. At each gate the system checks that the right evidence, cost limits and governance rules are satisfied before moving on.
Dynamic personas	Pre‑defined roles such as Data Engineer ($de), Data Scientist ($ds), Data Steward ($dst) etc. can be instantiated on‑the‑fly. Each persona has its own skill set and history, so the right expert handles each sub‑task.
Governance hooks	Built‑in safety policies stop dangerous Snowflake commands (e.g., DROP TABLE in prod) unless an explicit allow‑list overrides them.
CocoConsole	A read‑only browser‑based dashboard you launch with $cocoplus console. It shows flows, costs, session info, policy logs and other observability data.
CocoPilot	Optional natural‑language control – you can turn it on/off with $pilot on/off and give high‑level instructions in plain English.
CocoForge / Leviathan / Ronin	Higher‑level orchestration modes that let a team of agents work together autonomously on a goal, or run a “personal assistant” style loop that briefs you on progress.
CocoWisdom	A memory system that distils lessons from completed sessions, lets you query or reject knowledge, and keeps a searchable log of what the agents have learned.
CocoFlow	The engine that plans, scores, and dispatches work. It handles conflict detection, cost budgeting, checkpointing, and can pause for human approval at any stage.
CocoRoutine	Turns a finished flow into a Snowflake TASK so it can run on a schedule.

---

How you would use it

1. Install the plugin via the Snowflake Coco CLI (cortex plugin install …).
2. Initialize a new project with $pod init – this creates the required folder structure and AGENTS.md files.
3. Turn on the system: $cocoplus on.
4. Start the spec phase ($spec) to capture requirements. The system will spin up the appropriate persona agents (e.g., a Data Product Manager) to ask clarifying questions.
5. Progress through the phases ($plan, $build, $test, $review, $ship). At each gate the console shows you cost, evidence, and any policy violations.
6. Optional natural‑language orchestration – enable $pilot on and tell the system “optimise the ETL pipeline for cost”. The underlying agents translate that into concrete actions.
7. Review the wisdom – after a run you can query $wisdom get <topic> to see what the agents learned, or reject bad advice with $wisdom reject.
8. Schedule recurring work with $routine if the flow should run periodically.

---

Who would benefit?

Audience	Why it matters
Data engineering teams	Provides a repeatable, auditable pipeline for building and shipping Snowflake data products, with built‑in cost and safety checks.
Governance & compliance groups	Policy hooks automatically block risky operations and keep a log of every decision, making audits easier.
ML/AI platform teams	The same agentic framework can be extended to model‑training or feature‑store workflows, re‑using the same safety and observability stack.
DevOps / SRE	The scheduling (CocoRoutine) and monitoring (CocoConsole, $meter commands) give a familiar operational view of data‑pipeline health.

---

Limitations / things to keep in mind

• Coco‑only – It runs inside Snowflake’s Coco environment; you need the cortex CLI and a Snowflake account that supports Coco plugins.
• Complexity – The system defines many commands ($flow, $meter, $adversary, $diary, etc.). Teams will need a learning curve to adopt the full feature set.
• Safety policies are opt‑in – By default the autonomous reflection and some governance hooks are disabled; you must enable them deliberately.
• Open‑source, not officially supported by Snowflake – Bugs or missing features are handled by the community, not by Snowflake itself.

---

Quick start cheat‑sheet

# Install the plugin
cortex plugin install Snowflake-Labs/cocoplus
cortex plugin enable cocoplus

# Create a new project
$ pod init
$ cocoplus on

# Open the observability console (read‑only)
$ cocoplus console

# Run through the lifecycle
$ spec          # capture requirements
$ plan          # generate a plan
$ build         # execute ETL / model code
$ test          # run tests, cost checks
$ review        # human or automated review
$ ship          # deploy to production

# Optional natural‑language control
$ pilot on
# now you can type things like "reduce the cost of the nightly load by 20%"

# Schedule recurring runs
$ routine create my_daily_flow

---

License & legal

CocoPlus is released under the MIT license. It is open‑source, not an official Snowflake product, and carries the usual “as‑is” disclaimer.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.