Spawns swarms of agents coordinated through MCP tools, with shared memory, work claims and hive-mind consensus.
The summary describes what the project says it does; inclusion is not an endorsement.
Formerly Claude Flow; the packages still ship as claude-flow. Its own CLAUDE.md now says Ruflo records the work and Claude Code does it. Its issue tracker is an unusually detailed record of what goes wrong when many agents share one project.
github.com · Issue #70
“I noticed that when we have multiple agents or multiple Claude Code instances running, that it happens that they work on the same files sometimes resulting in overriding each other's work before changes are saved or committed.”
A first-person report with no trace attached. Closed in July 2025 with a maintainer comment announcing file locking in v2.0.0; no linked PR.
github.com · Issue #640, Verification Breakdown - The Root Cause
“Agents self-report "success" without mandatory verification”
A maintainer's warning about how one false claim of success passes through a chain of agents. Its three-agent example and 89% failure figure are illustrations, not a measured incident. Still open.
github.com · Issue #2028, What happened
“2 of 7 agents aborted before doing any work; 5 completed only because they ignored the brief and worked in "degraded mode" — reading source docs directly and using the shared memory namespace as a bus.”
Part of the cause is which tools Claude Code gives subagents, so not all of it is Ruflo. The fix was a documentation change in 3.7.0-alpha.69.
github.com · Issue #3178
“Under the repo's own "worktree-per-writer" rule, each worktree (and each Claude Code session with a different cwd) reads and writes a different claims.json, so claims_claim / claims_board / claims_steal do not coordinate the sessions they are meant to coordinate.”
Isolating the code also isolated the coordination. Shown by reading the source, not from an incident where two agents took the same task. The fix, PR #3210, was still unmerged in October 2026.
github.com · Issue #2633, Summary
“The result is that state is written to whichever subdirectory the session drifted into, and then never read again. Nothing errors.”
On one 12-repository workspace this left 109 state directories (97 stray) and about 25 daemons. Fixed in the reporter's fork; the upstream issue is still open. A 2025 issue (#765) describes the same root cause.
github.com · Issue #1637, Summary
“Persisted state after the background flush shows runCount=432. This creates the misleading impression that the CLI trigger did not execute, when in fact it did.”
Two writers for the same state make finished work look undone. Fixed by PR #1655 in v3.5.82.
github.com · Issue #765, Problem
“Agents are creating duplicate .claude-flow and .swarm folders in subdirectories instead of using the main workspace root folders.”
An early sign of sessions quietly no longer sharing memory. The v2 code it describes is gone, though the issue is still open.
github.com · Comment on issue #2621, The write drop is reproducible in ~30 lines
“Every one of the 50 writes returned success: true, and the DB stays integrity_check: ok. One process's entire batch was obliterated.”
Two processes wrote 25 rows each; all 50 writes were acknowledged and 25 rows remained. The test forced the sql.js fallback path, so it doesn't apply to every backend. The remaining race moved to #2878, closed in August 2026.
github.com · Issue #3931, Summary
“nextTask() would therefore hand a finished task to Claude again, which is a new billed turn.”
One word, "complete" instead of "completed", made five finished tasks look pending again. The issue shows the risk of the work running twice, not that it did. Fixed by PR #3932 and released in 3.56.2.
github.com · Issue #3290, Deep Dive Findings
“calculateRequiredVotes() derives the quorum threshold from state.workers.length, so a single caller could cross any strategy's threshold (raft/bft/quorum alike) by voting under fabricated identities — a Sybil attack on consensus.”
Shown in tests, not by an attacker or by agents colluding. Agent votes only mean something when identity is checked. Fixed by PR #3291; a follow-up (#3338, PR #3339) closed the same gap in sibling operations in October 2026.
github.com · Issue #3005, Impact
“Real risk of a contributor believing a fix is committed/shipped when their working tree has silently been reverted underneath them, especially if they don't diff before committing.”
With nine MCP processes and five agent worktrees running, a fix reverted twice with no git operation touching the file. The tooling itself became an unexpected writer. The cause is a hypothesis; the issue is still open.
github.com · Capability Brain and Governed Implementation
“Ruflo is the coordination ledger and policy decision point. Claude Code executes code, tests, commands, and file changes. A Ruflo coordination call records work; it does not perform the implementation.”
The project's own answer to the reports of swarms that did nothing. These are policies in a file, not proof they're enforced.
github.com · Issue #2584, Summary
“This is a data-durability bug in the AgentDB persistence layer: a long-running store should not be one interrupted/concurrent flush away from total loss.”
A shared memory database with about 16,700 entries became unreadable, then broke again about half an hour after a restore. PR #2585 added atomic flushes; cross-process locking was left for later.
github.com · Concurrent Automated Development
“Never allow two writers in one worktree. Give every writing agent an isolated worktree and explicit file ownership.”
The same section adds: "Only the integration owner edits shared manifests and lockfiles or reconciles overlapping changes."
github.com · Issue #1514, The Architectural Gap
“But the wire connecting agent_spawn → LLM provider → task execution is missing. This isn't a bug in one tool — it's a structural gap across the entire orchestration layer.”
An audit of v3.5.51. The maintainer fixed the hardcoded metrics it found (PR #1535) and disputed other points. The 99% headline is the auditor's claim, not a measurement.
github.com · Issue #2782, Relationship to #1637
“the maintainer's fix prevents cross-process corruption, but does not prevent two saveState() calls within the same process from racing on the shared <file>.tmp name.”
The fix for the run-count race introduced a new one: about 56,000 errors in one daemon log. The damage was log noise, not lost data. Fixed by PR #2788.
github.com · Issue #899, Bug Description
“Agents spawned via MCP tools (agent_spawn, mcp__claude-flow__agent_spawn) cannot communicate with each other using daa_communication.”
A spawned agent isn't automatically one other agents can reach. Reported against v2; triage in October 2026 says the subsystem no longer exists in v3.
github.com · Issue #3360, Summary
“When two ruflo MCP servers share one memory store, memory_retrieve on one server keeps returning a value after the other server has updated or deleted it.”
Shown with a two-process reproducer: a per-process cache served old values for up to five minutes while the database held the right one. Fixed by PR #3361 in September 2026.
github.com · Issue #1397, introduction
“The state-tracking fix from this issue was correctly applied — task_assign now transitions tasks to in_progress. However, agents never execute the assigned work.”
Seven agents reported as spawned, none did anything, and no file appeared. Closed as a duplicate of #1423, which PR #1435 fixed in March 2026. This shows v3.5.42, not later releases.
github.com · Discussion #696
“It has rules in its shared memory, and I keep asking it specifically to check the Claude.md file for the rules before it executes, but it just doesn't.”
Rules written in prose aren't enforced permissions. The reporter couldn't tell whether Ruflo or Claude Code was at fault, and no reply or log followed.
github.com · Issue #1748, What we found — per expectation (G5)
“Zero phases completed. Zero output.”
The swarm commands reported success on a three-phase task, while a single Claude Code subagent finished all three in about 45 seconds. The maintainer accepted the tool-wiring bug and merged fixes the same day.