Skip to content

Commit bbbdd69

Browse files
committed
Updated docs to account for what Anthropic just shipped.
1 parent 21e9903 commit bbbdd69

2 files changed

Lines changed: 69 additions & 14 deletions

File tree

README.md

Lines changed: 16 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,6 @@
11
# 🐝 Swarm Protocol
22

33
[![Tests](https://github.com/phuryn/swarm-protocol/actions/workflows/test.yml/badge.svg)](https://github.com/phuryn/swarm-protocol/actions/workflows/test.yml)
4-
![GitHub stars](https://img.shields.io/github/stars/phuryn/swarm-protocol?style=flat-square&color=0d9488)
54
[![License: MIT](https://img.shields.io/badge/License-MIT-0d9488?style=flat-square)](https://github.com/phuryn/swarm-protocol/blob/main/LICENSE)
65
[![PRs Welcome](https://img.shields.io/badge/PRs-welcome-0d9488?style=flat-square)](https://github.com/phuryn/swarm-protocol/blob/main/CONTRIBUTING.md)
76

@@ -36,6 +35,21 @@ This is a **coordination protocol, not a project management tool.** Jira was bui
3635

3736
The CLAUDE.md integration pattern is as important as the server itself. Agents coordinate automatically without humans configuring anything.
3837

38+
## How This Relates to Agent Teams, Code Review, and Other Multi-Agent Tools
39+
40+
These tools are complementary, not competing. They operate at different layers:
41+
42+
- **Claude Code Agent Teams** coordinates agents within a single session — one lead, parallel teammates, shared task list. Intra-session.
43+
- **Claude Code Review** dispatches parallel review agents against a single PR. Intra-PR.
44+
- **VS Code multi-agent** runs Claude, Codex, and Copilot agents side by side for one developer. Intra-machine.
45+
- **AgentConductor** dynamically adjusts agent communication graphs per coding problem. Intra-problem.
46+
47+
**Swarm Protocol** coordinates across sessions — multiple humans, each running their own agents, on a shared codebase. Inter-session, inter-human.
48+
49+
An agent uses Agent Teams internally for its own subtask parallelism. It uses Swarm Protocol's MCP tools to know which work to pick up, which files to avoid, and what context the previous agent left behind.
50+
51+
The single-player problem is getting solved. The multiplayer problem hasn't started.
52+
3953
## Who Is This For
4054

4155
Teams of 2+ developers where AI agents (Claude Code, etc.) are the primary development interface. If your workflow is "open terminal → tell Claude Code what to build → review the PR" and teammates are doing the same thing at the same time — this is the missing layer.
@@ -63,7 +77,7 @@ See [LANDSCAPE.md](docs/LANDSCAPE.md) for the full competitive breakdown.
6377
| **Intent** | A unit of desired outcome — not a ticket. Lifecycle: `draft → open → claimed → done`. Has constraints, acceptance criteria, and dependency chains. |
6478
| **Claim** | "I'm working on this." Tracks which files are being touched. Includes heartbeat — claims with no heartbeat for 30 min get flagged as stale. |
6579
| **Signal** | Event notification: completion, blocked, conflict, info. When a completion signal fires, dependent intents auto-unblock. |
66-
| **Context Package** | Everything an agent needs to start work — intent, dependencies, active claims on overlapping files, recent signals, team conventions — assembled in one call via `get_context`. |
80+
| **Context Package** | Everything an agent needs to start work — intent, dependencies, active claims on overlapping files, recent signals, team conventions — assembled in one call via `get_context`. Solves the state handoff problem: when Agent A completes work, the structured output travels with the task so Agent B gets a stable input contract, not just "the file changed." Sequential dependency chains require this — Agent B's correctness depends on Agent A's output schema being stable across runs. |
6781

6882
## Architecture
6983

docs/LANDSCAPE.md

Lines changed: 53 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -12,11 +12,15 @@ This is single-player multiplayer vs. true multiplayer. Different product catego
1212

1313
## Existing Tools (All Single-Developer Focus)
1414

15-
### 1. Claude Code Agent Teams (Anthropic, experimental)
16-
- One lead agent coordinates teammates within a single session
17-
- Shared task list, file locking, inter-agent messaging
15+
### 1. Claude Code Agent Teams (Anthropic, shipped with Opus 4.6)
16+
- Officially launched as experimental feature — enable with `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS`
17+
- One lead agent coordinates teammates, each in independent context windows
18+
- Shared task list, inter-agent messaging, peer-to-peer communication between teammates
19+
- Teammates work independently — not subagents reporting back to a parent
1820
- Intra-session only — not cross-team, not cross-human
19-
- Still experimental, known limitations around session resumption
21+
- Known limitations: no session resumption, no nested teams, no cross-session state
22+
- Uses significantly more tokens than single sessions (~3-4x for a 3-teammate team)
23+
- Best for: research/review, new modules, debugging with competing hypotheses, cross-layer coordination
2024
- https://code.claude.com/docs/en/agent-teams
2125

2226
### 2. Claude Code Tasks (Anthropic, native)
@@ -26,14 +30,34 @@ This is single-player multiplayer vs. true multiplayer. Different product catego
2630
- Dependencies and blockers supported
2731
- https://venturebeat.com/orchestration/claude-codes-tasks-update-lets-agents-work-longer-and-coordinate-across
2832

29-
### 3. CCPM — Claude Code Project Manager (automazeio)
33+
### 3. Claude Code Review (Anthropic, March 2026)
34+
- Multi-agent PR review — dispatches parallel agents examining code from different perspectives
35+
- Verification agent aggregates findings, removes duplicates, ranks by severity
36+
- Agents don't approve PRs — they post flagged issues as inline comments
37+
- Configurable via REVIEW.md (what to prioritize/deprioritize) and CLAUDE.md (codebase context)
38+
- 54% of PRs receive substantive comments (up from 16% with older approaches per Anthropic)
39+
- Scoped to single PR — not cross-session, not cross-team coordination
40+
- Team and Enterprise plans only, ~$15-25 per review, ~20 min completion time
41+
- https://www.anthropic.com/research/claude-code-review
42+
43+
### 4. VS Code Multi-Agent Development (Microsoft, Feb 2026)
44+
- VS Code 1.109 positioned itself as "the home for multi-agent development"
45+
- Run Claude, Codex, and Copilot agents side by side — local, background, or cloud
46+
- Parallel subagents with visibility into what each is doing
47+
- Agent Sessions view: single place to manage all running agents
48+
- Agent Skills (Anthropic's open standard) generally available in VS Code extensions
49+
- Custom agents with specialized tools, instructions, and model selection per agent
50+
- Still single-developer scope — multiple agents for one person, not team coordination
51+
- https://code.visualstudio.com/blogs/2026/02/05/multi-agent-development
52+
53+
### 5. CCPM — Claude Code Project Manager (automazeio)
3054
- GitHub Issues as database, git worktrees for parallel execution
3155
- PRD → Epic → Task → Issue → Code → Commit pipeline
3256
- Closer to our idea but assumes single developer dispatching agents
3357
- No cross-team coordination or conflict detection
3458
- https://github.com/automazeio/ccpm
3559

36-
### 4. tick-md (Purple Horizons)
60+
### 6. tick-md (Purple Horizons)
3761
- Markdown file as database, git-backed, MCP server included
3862
- 7 GitHub stars — tiny traction
3963
- File locking, dependency tracking, real-time monitoring
@@ -42,41 +66,41 @@ This is single-player multiplayer vs. true multiplayer. Different product catego
4266
- https://www.tick.md/
4367
- https://github.com/Purple-Horizons (org page)
4468

45-
### 5. Agent-MCP (rinadelph)
69+
### 7. Agent-MCP (rinadelph)
4670
- Shared memory graph ("Obsidian for AI agents"), file locking, task management
4771
- Real-time dashboard visualization
4872
- Single-repo, single-developer scope
4973
- https://github.com/rinadelph/Agent-MCP
5074

51-
### 6. 1Code (21st.dev, YC W26)
75+
### 8. 1Code (21st.dev, YC W26)
5276
- GUI wrapper for Claude Code / OpenAI Codex
5377
- Cloud-based background execution, kanban board
5478
- Git worktree isolation per agent
5579
- GitHub/Linear/Slack automation triggers
5680
- Orchestration client, not a coordination protocol
5781
- https://dev.to/_46ea277e677b888e0cd13/1code-managing-multiple-ai-coding-agents-without-terminal-hell-14o4
5882

59-
### 7. multi-agent-coordination-mcp (AndrewDavidRivers)
83+
### 9. multi-agent-coordination-mcp (AndrewDavidRivers)
6084
- MCP server for Cursor IDE specifically
6185
- File locking, dependency management, Projects → Tasks → Todo Items
6286
- Single-developer scope
6387
- https://github.com/AndrewDavidRivers/multi-agent-coordination-mcp
6488

65-
### 8. GitButler
89+
### 10. GitButler
6690
- Auto-sorts parallel Claude Code sessions into separate git branches via hooks
6791
- Each session gets its own branch automatically
6892
- Smart but solves git conflict isolation, not team coordination
6993
- https://blog.gitbutler.com/parallel-claude-code
7094

71-
### 9. multi-agent-coordination-framework (timothyjrainwater-lab)
95+
### 11. multi-agent-coordination-framework (timothyjrainwater-lab)
7296
- Methodology/protocol repo, not a tool
7397
- Built by a non-technical operator coordinating Claude + GPT agents
7498
- Handoff checklists, consistency gates, structured memo formats
7599
- "If it's not in a file, it doesn't exist" — core principle
76100
- Proves the pain is real but solution is manual protocols
77101
- https://github.com/timothyjrainwater-lab/multi-agent-coordination-framework
78102

79-
### 10. Beads Village (MCP-based)
103+
### 12. Beads Village (MCP-based)
80104
- MCP server with task queues, file locking, and built-in messaging between agents
81105
- Standard workflow: initialize → claim tasks → lock files → work → complete
82106
- Runs entirely locally, data stored in Git repo
@@ -86,6 +110,20 @@ This is single-player multiplayer vs. true multiplayer. Different product catego
86110

87111
---
88112

113+
## Research Papers
114+
115+
### AgentConductor (Chinese labs, Feb 2026)
116+
- Dynamic topology evolution for multi-agent code generation
117+
- LLM-based orchestrator infers agent roles and task difficulty, constructs task-adapted DAG topology
118+
- Easy tasks get small, cheap teams. Hard tasks get large, highly connected teams.
119+
- Manager rewrites the team workflow on failure based on error feedback
120+
- 68% token cost reduction vs. fixed topologies, 14.6% accuracy improvement on competition-level code
121+
- Validates our core thesis: static agent pipelines waste compute on simple tasks and fail on complex ones
122+
- Still single-problem scope — solves intra-session topology, not cross-session coordination
123+
- Paper: https://arxiv.org/abs/2602.17100
124+
125+
---
126+
89127
## What Jira Is Doing
90128

91129
- Atlassian launched **"agents in Jira"** (open beta, Feb 2026) — assign tasks to AI agents from the same dashboard as human employees
@@ -137,11 +175,14 @@ This is single-player multiplayer vs. true multiplayer. Different product catego
137175
| Team support | None | Multi-team with conventions |
138176
| State sharing | Env vars, shared files | PostgreSQL, real-time via MCP |
139177
| Draft workflow | No | Draft → publish → claim |
178+
| State handoff | Implicit (file system changes) | Explicit (Context Package with structured output + dependency schema) |
140179

141180
## Key Positioning
142181

143182
- Not "Jira but with AI" — that's what Atlassian is doing and they'll always do it better
144183
- Not "one dev managing agent fleet" — that's what 1Code, CCPM, tick-md do
184+
- Not "smarter agents in one session" — that's what Agent Teams, Code Review, VS Code multi-agent, and AgentConductor do
145185
- **"Coordination infrastructure for agent-first teams"** — a new category
146186
- The market for this is tiny today and enormous in 12-18 months
187+
- March 2026 signal: Anthropic shipped Agent Teams AND Code Review. Microsoft shipped multi-agent VS Code. AgentConductor paper dropped. All single-player. The multiplayer gap is widening, not closing.
147188
- First mover who names the category owns the narrative

0 commit comments

Comments
 (0)