TJP

Tyler J. Peckenpaugh / Agentic development

Agent tooling landscape

The agent tools I looked at for The control plane got crowded, sorted into the planes from the tooling anatomy. Each entry links to the source it's based on. As of November 12, 2026.

A working interpretation, not a ranking. Plane assignments are my judgment about where most of a tool's work lands. T3 Code, Orca and Gas Town were read from source without being run; the rest come from vendor pages and repositories, which change quickly.

Tools by plane

ToolKindPlanesWhat it does
Cursor 3Agent IDEControlDeliveryRebuilt in April 2026 around a workspace for running local and cloud agents side by side.
Google Antigravity 2.0Agent IDEControlDeliveryMoved agent management into a standalone desktop app, separate from the editor, in May 2026.
Devin DesktopAgent IDEControlDeliveryFormerly Windsurf. Opens to an Agent Command Center, a board of local and cloud agents, and runs ACP agents.
JetBrains AirAgent IDEControlDeliveryAn agentic development environment for delegating coding tasks to agents; public preview from March 2026.
WarpAGPL-3.0 (client)TerminalControlDeliveryCalls itself an agentic development environment; its client was open-sourced in April 2026.
GitHub Agent HQPlatformControlDeliveryMission Control for Copilot, Claude and Codex sessions, tied to issues and pull requests.
Agent Client ProtocolApache-2.0ProtocolControlStarted by Zed so one interface can drive many agents; has an agent registry and support from JetBrains and others.
T3 CodeMITAgent workspaceControlDeliveryA server with web, desktop and mobile clients that drives the agents you already have through their own protocols, with worktrees on request. About 57,000 stars.
OrcaMITAgent workspaceControlDeliveryDesktop app with a worktree per task, status hooks for each agent, and a CLI that lets agents start and message other agents. About 122,000 stars.
NimbalystMITAgent workspaceControlDeliveryFormerly Crystal. Runs agents in parallel and lets you edit their work visually in markdown, mockups and diagrams, and track tasks.
ConductorAgent workspaceControlDeliveryMac app for running agents in worktrees. Its site now says it is joining Meta.
Vibe KanbanApache-2.0Agent boardControlDeliveryLocal kanban board that runs an agent per card. Bloop shut down in April 2026 and the project is now community maintained.
KandevAGPL-3.0Agent boardControlDeliverySelf-hosted kanban for orchestrating agents, reviewing changes and opening pull requests.
Claude SquadAGPL-3.0TerminalControlManages several terminal agents, each in its own worktree.
Gas TownMITAgent organizationControlDeliveryNamed roles: a mayor assigns work, a witness checks on stuck workers, a refinery runs a merge queue. Built on Beads.
BeadsMITShared stateControlIssue tracker for agents. Data now lives in a versioned database under its own Git ref.
PaperclipMITAgent organizationControlOpen-source orchestration for teams of agents: org charts, budgets, tickets and scheduled check-ins.
MCP Agent MailShared stateControlMessages between agents plus advisory file reservations with an expiry and an optional pre-commit guard. Deprecated in favor of a Rust rewrite, mcp_agent_mail_rust.
KiroSpec-driven IDEKnowledgeDeliveryEach spec produces requirements, design and task files before implementation.
Spec KitMITSpec frameworkKnowledgeDeliveryCommands that take a feature from specification to technical plan to tasks.
OpenSpecMITSpec frameworkKnowledgeSpec-driven development for AI coding assistants, with specs kept in the repository.
BMAD-METHODMITSpec frameworkKnowledgeDeliveryBriefs, specifications and architecture that it says keep decisions explicit and carry context forward.
Augment IntentAgent workspaceKnowledgeControlA coordinator agent writes a spec for each task that every agent in the workspace works from.
Backlog.mdMITTask trackerKnowledgeControlEach task is a markdown file in the repository, with a CLI, an MCP server and a local kanban web UI.
SuperpowersMITSkillsKnowledgeDeliverySkills for brainstorming, planning, test-driven development, review and verification.

Star counts are from GitHub on November 12, 2026.

How policy reaches workers

How each of the three tools I read passes a project's instructions to the agents it starts, in the mechanism terms from the tooling anatomy.

ToolHowMechanism
T3 CodeLeaves it to the agent. When it starts Claude Code through the SDK, it asks it to load project settings and instruction files.Harness loads instruction files
OrcaRuns the agent's own CLI in the worktree, so it reads the same instruction files. Its orchestration guide also tells the coordinator to put constraints in each worker's task spec.Harness loads instruction files; agent-directed prose
Gas TownA session-start hook runs gt prime, which injects the role's instructions plus operator-written role directives that override the workflow where they conflict.Harness hook injects role policy

Where shared state could live

Options for recording which agent has claimed which task and which files it is changing, so agents in other worktrees can see it. This is an open question, not a recommendation.

WhereExampleVisible toOwned by
The tool's own databaseOrca, T3 Code (SQLite and JSON in your home directory)Every agent that tool startsThe tool
Files committed to the repositoryBeads' original designOther branches only after mergingThe project
A coordination serverMCP Agent Mail's file reservationsEvery agent connected to the serverThe server
A versioned database under a Git refBeads today, using DoltEvery worktree; other machines by push and fetchThe tool, stored in the repository
The shared .git directory or custom Git refsBeads' database uses a ref; I found no tool keeping simple claim records this wayEvery worktree; refs also by push and fetchUndecided