Skip to main content

10. Competitive Benchmark

State: Research

Source review date: 2026-08-19

Purpose And Evidence Boundary​

This is a dated, source-inspected comparison of Seraph on develop with OpenClaw, Hermes Agent, and IronClaw. It identifies product pressure; it does not define Seraph's target, execution order, or shipped status. Those belong to the Project Constitution, the GitHub Project, and Development Status.

Competitor entries are based on official documentation, repositories, parity matrices, and releases. They prove what a project currently documents or ships in source; they are not independent reliability or quality benchmarks. No overall parity or superiority claim is allowed from this table.

Evidence labels used here:

  • Documented: described by the project's official material.
  • Source-inspected: an official implementation, release, or self-reported parity boundary was inspected.
  • Runtime-verified: exercised against a running reference system by Seraph's maintainers.
  • Independently benchmarked: compared under the same published task and scoring protocol.

This refresh is Source-inspected; it is not Runtime-verified or Independently benchmarked.

How To Read The Matrix​

  • Ahead: comparable evidence shows Seraph's shipped surface is stronger.
  • Behind: the competitor's documented/source-inspected surface is materially more complete than Seraph's shipped surface.
  • At Par: comparable evidence shows no material difference.
  • Unknown: evidence or comparability is insufficient.

Unknown is the default when one side has only a deterministic receipt, marketing claim, or differently scoped implementation.

Current Axis Matrix​

AxisOpenClawHermesIronClawSeraph evidence boundary
Persistent goals and bounded follow-throughBehindBehindUnknownSeraph has goals, strategist ticks, proactive delivery, and feedback foundations, but no shipped goal -> capability job -> verified outcome -> changed-decision loop.
Durable jobs, flows, and restart recoveryBehindBehindUnknownDurable workflow foundations exist, but the universal capability job/checkpoint contract remains Partial.
Scheduled and event-driven proactivityBehindBehindUnknownScheduled work exists; production standing-intent admission, execution, and outcome closure remain incomplete.
Typed execution and isolationUnknownUnknownUnknownAll four projects document different boundaries and caveats; no shared escape/exfiltration benchmark establishes a ranking.
Memory governance and outcome learningUnknownUnknownUnknownCanonical memory and guardian learning foundations exist, but comparative recall quality and behavior change are not independently benchmarked.
Operator visibility and recoveryUnknownUnknownUnknownSeraph has a dense cockpit, but task efficiency and recovery burden have no current cross-system benchmark.
Browser and external-system executionBehindBehindUnknownSeraph's catalog and proof surfaces do not substitute for broad live task completion.
Cross-surface reachBehindBehindBehindThe authenticated paired Mac edge and canonical cross-surface identity are not shipped.
Guardian restraint and intervention qualityUnknownUnknownUnknownSeraph has differentiated scaffolding, but no comparable live outcome study proves an advantage.
One-GPU priority and non-starvationUnknownUnknownUnknownSeraph's serial broker is an accepted target, not a shipped cross-consumer receipt.

Every Behind cell is scoped only to documented surface completeness where Seraph's own status marks the corresponding end-to-end behavior Partial or missing. It does not score reliability, security quality, task success, or operator outcomes. Those dimensions remain Unknown until the systems are run under a shared protocol.

Reference-System Pressure​

Hermes Agent: goal completion and operator velocity​

Hermes documents direct pressure on Seraph's goal-directed execution gap. Persistent Goals continue work until completion, pause, or a budget boundary. Kanban adds durable tasks, dependencies, reclaim, human-unblock state, and structured handoffs. Cron, delegation, browser/file/terminal tools, skills, memory providers, and messaging make those loops immediately usable from a compact operator surface.

Hermes also documents important limits. Delegated children are isolated but do not provide universal crash-resumable child execution, and its security guide warns that host terminal access is not an adversarial sandbox. Seraph should copy the clarity and completeness of the goal loop, not weaken its own trust boundary to match feature count.

Official sources reviewed:

OpenClaw: durable automation, reach, and memory governance​

OpenClaw documents a broad operating surface: a gateway, Control UI, channels, paired nodes, browser execution, plugins, skills, multi-agent routing, heartbeats, durable automations, standing orders, and typed resumable flows. Lobster and Task Flow are particularly relevant because they combine bounded multi-step execution, approval checkpoints, resume state, timeouts, and operator inspection.

Its own security material also bounds the comparison: OpenClaw is designed around a trusted single-operator model, sandboxing is configurable rather than universal, and its public extension ecosystem creates supply-chain pressure. Seraph should learn from the control-plane coherence while retaining stricter canonical memory and execution authority.

Official sources reviewed:

IronClaw: explicit execution boundaries​

IronClaw provides an explicit typed execution-boundary reference: capability permissions, WASM isolation, mediated credentials, endpoint allowlists, leak scanning, resource limits, durable database state, and distinct authorization, approval, reservation, dispatch, execution, and evidence phases.

Its current breadth should not be overstated. The project's own parity matrix, self-dated as last reviewed against OpenClaw on 2026-05-02, marks parts of browser automation, active memory, multi-agent operation, and other OpenClaw-compatible behavior incomplete. IronClaw is therefore a useful security-contract reference, not proof that every advertised agent loop is more mature than Seraph's.

Official sources reviewed:

Capability-First Implications For Seraph​

The highest-leverage gap is not another capability catalog or proof endpoint. It is one complete proactive operator journey:

goal or standing intent
  -> candidate intervention
  -> deterministic admission and priority
  -> approval / reservation
  -> typed capability job
  -> real bounded execution
  -> evidence and external readback
  -> outcome evaluation
  -> governed canonical-memory update
  -> changed later decision

Research implications for tracked work:

  1. Make the universal goal-conditioned loop the first end-to-end capability slice, while preserving durable jobs, global GPU arbitration, and secure execution as hard dependencies.
  2. Measure restart recovery, cancellation, approval-to-readback correctness, false triggers, no-invent-work behavior, and changed later decisions.
  3. Treat memory as a moat only when provenance, confidence, supersession, expiry, contradiction handling, false promotion, and action boundaries are measurable.
  4. Prefer one authenticated identity across cockpit, paired edge, voice, and messaging over a race for raw channel count.
  5. Require extensions to support provenance, staged install, update, disable, revoke, conformance checks, and rollback before ecosystem breadth counts as product progress.
  6. Publish axis-specific comparable receipts before changing any claim from Unknown or Behind to At Par or Ahead.

The Strategy Claim Ledger remains the wording gate. The Agent Competition Truth Table provides historical benchmark context and should not override this later dated source refresh.