Sherpa

The framework

A collaboration framework for agentic workflows with a focus on governance, research, and execution. Multi-project federation, behavioral agent definitions, a 9-backend dispatch pipeline, semantic knowledge indexing, and an initiative lifecycle with MCP-exposed governance — infrastructure that lives in your codebase and versions with git.

Seven pillars

Each pillar is a working system with real data behind it — not a roadmap item.

Behavioral Agent System

Agent roles define what an agent does — its defaults, quality bar, and fail triggers — not who it pretends to be. That keeps behavior predictable and reviewable. 12 roles run in production this way, each scoped to the task types it can take, an approach validated against Anthropic and academic research.

Governance Engine

Every piece of work moves through proposal → review → integration, so nothing ships without a decision trail. 73 real initiatives have run through this lifecycle. External agents participate through 7 policy-gated MCP tools.

Execution Pipeline

Describe a task once and the pipeline routes it to the right agent and backend, logs every event, and brings the result back for review. 9 dispatch backends — 5 CLI, 3 API, 1 remote gateway. Authority leases with fencing tokens mean two agents can never write the same artifact at once.

Studio Application

One dashboard shows the whole system — task boards, initiative trees, research feeds, and workforce views over live governance data. 25 routes and 93 React components, with one Studio instance able to govern multiple codebases. Deploys blue-green with zero downtime.

Executable Conventions

Conventions here are protocols agents run, not documents people forget. 14 skills define step-by-step workflows for research, review, and planning. 8 rules load automatically by file pattern, and provenance tracking records who wrote each document and when it was last verified.

Config-as-Code

One config file declares how your team works — dispatch routing, governance policy, knowledge backends, vocabulary — and versions with git like everything else. Three directories keep concerns separate: .sherpa/ for tool state, .claude/ for agent config, docs/ for human prose. Type-safe with JSON schema.

Convention Sync CLI

Any project can adopt the framework in one command. sherpa init scaffolds the governance structure; sherpa sync keeps conventions current as they evolve. Provenance frontmatter tracks authorship, review state, and verification date for every maintained document.

Cross-cutting infrastructure

Capabilities that span multiple pillars and make the framework more than conventions in a filesystem.

Multi-Project Federation

Run one governance substrate across many codebases and see them all in one place. A project registry merges task boards, research feeds, and initiative trees into aggregate views, so a team adopting Sherpa on five repos still has one place to look.

Semantic Knowledge Engine

Search and cluster everything the framework has produced — without standing up a vector database. The default backend is zero-dependency and algorithmic (TF-IDF indexing, extractive summaries, agglomerative clustering); swap in another backend when you need one. Queryable over MCP.

SQLite State Layer

Markdown stays the source of truth; SQLite gives you fast, queryable state derived from it. Three embedded databases handle coordination (authority leases, sessions), events (the audit trail), and the knowledge index. WAL mode keeps concurrent reads safe.

MCP Coordination

Multiple agents can work the same project at once without stepping on each other. An HTTP Streamable transport and session manager enforce authority leases, and fencing tokens prevent concurrent mutation of the same artifact. Governance applies only to autonomous agents — it never constrains a human+AI session.

Packages

@sherpa/studio-coreDomain logic, config schema, lifecycle engine, dispatch routing, knowledge backend interface
@sherpa/studio-ui93 React components for governance visualization
@sherpa/studio-mcpMCP server with authority leases, session management, and initiative governance tools
@sherpa/studio-cliConvention scaffolding, sync, and provenance tracking

This is Sherpa governing its own development.

Screenshots from Sherpa Studio — the dashboard this repository runs on, with its real governance data.

Sherpa Studio process workspace showing an initiative tree with lifecycle states and a selected initiative's review trail
73 initiatives tracked from proposal to integration — every node is a real decision with a review trail.
Sherpa Studio task board listing dispatched agent missions with per-backend filter chips
100 dispatched agent missions across 9 backends. 38 shipped.
Sherpa Studio research dashboard showing a heartbeat schedule and dated reports grouped into intelligence streams
Autonomous research running on a heartbeat, 8am–11pm, feeding nine intelligence streams.

Built by Rob Abby

This framework emerged from building agentic workflows and discovering the governance layer didn't exist. Every convention and agent role here is in production, governing this framework's own development. Source is private while open-source prep completes — happy to walk through it.