Each system solves one problem end to end. Some pairs are designed to complement each other: one owns what the other deliberately leaves out. They connect only through public interfaces, so you can adopt one, both, or swap either for something else.
Complementary products tend to drift into mutual dependency. These rules keep the benefit of combining them without making either one a prerequisite.
01
Each installs alone
No default install requires its partner. Every product is useful on its first day without the other.
02
Public interfaces only
Partners call each other's documented CLI or adapter contract. Neither reads the other's private files or state.
03
One owner per fact
Each piece of data or decision has exactly one home. Nothing is copied into the partner "just in case".
04
Degrade visibly
When a partner is missing or broken, the product falls back or reports degraded. It never silently substitutes or invents a result.
+
Holoself + Nextstep
Who you are, and what you decided.
Available today
Holoself keeps your approved, reusable self: identity, voice, stories, and evidence. Nextstep keeps the record of one job search: opportunities, decisions, submissions, and outreach. Together, your agent drafts from your real evidence while every decision lands in an auditable record, and neither product holds the other's data.
Ownership never moves: Holoself doesn't store applications, and Nextstep doesn't copy your story bank.
Three ways to run it
Pick the one that fits. You can move between them later.
Standalone
Nextstep alone
The full job-search record: decisions, submissions, outreach, strategies.
A small native candidate-profile card supplies the essentials about you.
AgentFlow governs delivery: approved intent, handovers, evidence, and acceptance. Meshloop executes engineering work: task graphs, isolated worktrees, checks, and repair. Paired through an optional provider adapter, AgentFlow can hand a scoped build to Meshloop and judge what comes back, while Meshloop's "pass" is never the final word.
The dashed Build step is the only point of delegation. Scope flows down, evidence flows up, and acceptance stays with AgentFlow.
Three ways to run it
Both products work today on their own. The pairing is being qualified.
Standalone
AgentFlow alone
Roles, handovers, deterministic gates, and acceptance for AI-assisted delivery.
Executes through Claude Code, Codex, Agy, Pi, Grok, or an explicit manual path.
No Meshloop, collector, or other product required.
Best when you need accountable delivery with the tools you have. Explore AgentFlow →
Standalone
Meshloop alone
Its own human gates: review the plan, accept, then integrate.
Runs from the CLI, slash commands, or MCP without any AgentFlow configuration.
Isolated worktrees and compiler-verified repair on day one.
Best when you want parallel agents with safe isolation now. Explore Meshloop →
Better together · preview
Both
AgentFlow's accountability around Meshloop's parallel execution.
Meshloop works inside an approved scope and narrowed limits it cannot widen.
Evidence returns digest-verified and bound to the tested candidate.
An optional Meshloop provider is on AgentFlow's development branch. It is not in a release yet.
Guarantees designed into the boundary
From AgentFlow's engineering-provider profile, which is proposed and still under qualification.
Scope can't growMeshloop's task graph can refine the approved scope but never enlarge it.
A pass is not acceptanceA technical receipt is evidence. AgentFlow validates it independently before accepting.
Unknown stays unknownA timeout or lost reply is reconciled, never recorded as a fabricated pass.
Mismatch fails earlyAn unsupported version or capability blocks before any work is dispatched.
No shared internalsAgentFlow never reads Meshloop's database; Meshloop never imports AgentFlow's contracts.
Independent upgradesEach product upgrades on its own within a tested compatibility matrix.