AgentsAgent orchestrationWorkflow systemsAgent workflow orchestration

01 — Workflow-orchestration prior art

Compares local orchestrators, coding-agent workbenches, workflow engines, control protocols, and shared-memory systems.

Summary

The market strongly validates visible execution, durable task state, provider mixing, review gates, and issue or PR visibility. It does not yet supply a drop-in runtime that preserves jackin Capsule sessions, isolation, and host-side policy.

Question and scope

Which external systems should jackin reuse, adapt, benchmark, or avoid when designing agent workflows?

The scope excludes SaaS-only systems as runtime dependencies. Commercial products remain useful UX evidence.

Method

Projects were compared by execution substrate, session visibility, branch/worktree ownership, workflow semantics, human gates, GitHub lifecycle, provider breadth, local-first operation, and compatibility with jackin host-mutation rules. Maturity and license require revalidation before dependency use.

Findings

Workflow engines and coding-agent orchestrators

ProjectUseful evidenceCurrent stance
ConductorDeterministic YAML graphs, first-match routing, scripts, parallelism, human gates, dry runsGraph-semantics reference; possible adapter after stable jackin APIs
ContrabassIssue-driven plan/execute/verify flow, worktrees, worker livenessHands-on benchmark; duplicates execution surfaces
agtxLocal board, artifacts, phase transitions, MCP, multi-agent pluginsTask/artifact reference or MCP peer
OptioTicket-to-PR pipeline, self-healing CI/review, queue, dashboardArchitecture benchmark; too infrastructure-heavy for the local substrate
SandcastleLibrary runner, sandbox providers, branch strategies, hooks, logsReuse candidate only if host hooks and cross-language boundary fit policy
OpenClaw Code AgentPlan approval, persistence, suspend/resume/fork, explicit goal loops, action buttonsIntervention-state reference
SwitchboardRecurring local agent workflowsLater queue/daemon reference
Ruah and BernsteinDAGs, file claims, merge ordering, audit and governanceResearch references pending source/license audit
RuFloBroad swarm/MCP surface and dashboard demandAvoid as dependency without a narrow verified feature

Local workbenches and execution substrates

ProjectUseful evidenceCurrent stance
Vibe KanbanLocal task board, worktree-per-task, diff review, alerts, agent breadthBenchmark before building a competing board
Handler.devContainer/Firecracker canvas, terminal per sandbox, workspace forksIsolation and forkable-workspace benchmark
HelmorDesktop planning, execution, review, test, merge, parallel comparisonUX benchmark
OrcaWorktree, terminal, browser, and diff per taskUX benchmark for parallel visible sessions
SculptorVisible agents in isolated Docker containers, pairing syncClosest substrate comparison; possible future adapter
EmdashIssue intake, worktree, agent, diff, PR, CI, merge gates, broad agent supportStrongest roadmap-to-pr hands-on benchmark
workmux, worktrunk, parWorktree/session naming and cleanup ergonomicsLow-level references only

Review, terminal, and protocol surfaces

ProjectUseful evidenceCurrent stance
MCO / HiveMulti-provider review, consensus, deduplication, SARIF/Markdown, ACPOptional review adapter after branch state is authoritative
Forge MCP and terminal-mcpPersistent PTY create/read/wait/send APIsAPI-shape reference; jackin should back its own surface with Capsule
ACPStructured JSON-RPC agent transportWatch and interoperate; do not invent an incompatible provider transport
GitHub Agentic Workflows and third-party agentsPermission narrowing, safe outputs, GitHub-native issues/PRs/commentsVisibility/security reference, not local execution substrate

Shared memory

ProjectUseful evidenceCurrent stance
OpenMemory / Mem0Project-scoped memory over MCPBackend/API-shape reference
Letta shared memoryShared blocks, read-only policy, concurrent updatesStrong model for workflow-scoped briefs
Zep / GraphitiTemporal knowledge-graph memoryFuture retrieval candidate after simpler local memory proves useful

Dependency verdict

No reviewed project currently qualifies as the authoritative workflow runtime dependency. MCO/Hive is the clearest optional review adapter. Conductor and ACP are plausible future interoperability surfaces. Workbenches should be benchmarked because they may own the operator UI better, while jackin remains the isolation/session provider.

A dependency verdict should change if an external orchestrator provides a stable interface for externally supplied sandbox/session providers, preserves Capsule-owned visibility, avoids silent host mutation, and supports typed gates and durable evidence.

Implications for jackin

  • Keep the run/session API substrate-neutral so workbenches can drive jackin later.
  • Borrow deterministic graph, gate, artifact, and intervention concepts without inheriting host-native tmux/worktree ownership.
  • Benchmark current workbenches before building a dashboard or task board.
  • Require a concrete, narrow feature plus license/security review before adding an orchestration dependency.

Limitations and unknowns

This is a source-and-product-surface comparison, not a uniform hands-on benchmark. Project maturity, licenses, APIs, and commercial availability can change quickly.

Sources

Primary project and product links are embedded in the comparison tables.

On this page