18. Agent Competition Truth Table

Purpose

This is the broad M0 competition truth table for Seraph's capability-first world-class strategy.

As of: 2026-05-04

This document covers source-backed pressure from major agent systems and frameworks. It does not make superiority claims for Seraph. It records what public official/source evidence says competitors are strong at, how confident we are in that read, and which Seraph milestones must answer the gap.

Use this with:

• 17. Seraph World-Class Strategy
• 19. Strategy Claim Ledger

Confidence Legend

Confidence	Meaning
High	Current official docs, product docs, or primary repository evidence directly supports the row.
Medium	Official/source evidence exists, but the public surface is narrower, changing, deprecated, or incomplete for some axes.
Low	Current primary evidence is insufficient for a strong competitive claim; use only as a watch item.

Milestone Key

Milestone	Strategic gap class
M0	Competition truth, claim discipline, source hygiene, and benchmark governance
M1	Capability kernel, manifests, taxonomy, and contribution contracts
M2	Execution supremacy across terminal, process, files, browser/computer use, patching, artifacts, and repair
M3	Trusted execution boundaries, isolation, credentials, approvals, prompt-injection defense, and provider trust
M4	Selective reach through native, browser, messaging, node, webhook, and external channels
M5	Jobs, routines, workflows, delegation, checkpoint, branch, resume, compare, and background ownership
M6	Memory superiority with provenance, confidence, freshness, privacy, correction, and behavior-changing recall
M7	Dense cockpit and activity ledger for inspection, approval, routing, spend, artifacts, failures, and recovery
M8	Guardian brain over the capability substrate: salience, timing, restraint, goals, feedback, and follow-through
M9	Governed ecosystem, managed connectors, package trust, versioning, compatibility, and review flows

Truth Table

System	Current source-backed shape	Primary/source URLs	Confidence	Main Seraph pressure	Gap map
Hermes Agent	Dense operator agent with broad built-in tools, toolsets, MCP, terminal/process/files, browser automation, memory, delegation, cron/background work, messaging, skills, and defense-in-depth security controls.	Docs, tools, browser, MCP, security	High	Seraph must match raw capability breadth without losing guardian governance, receipts, and cockpit legibility.	M1, M2, M3, M4, M5, M7, M9
OpenClaw	Gateway-centered agent/control plane with Control UI, browser execution, tools/plugins/skills, multi-agent routing, node/device pairing, channel reach, and gateway security posture.	Docs, Control UI, browser, architecture, security	High	Seraph needs at least comparable control-plane legibility, channel governance, browser safety, and node/channel reach, but with guardian memory and stricter supervision.	M1, M2, M3, M4, M5, M7, M9
IronClaw	Security-first agent OS/reference with public source framing around capability permissions, sandboxing, credential boundaries, routines/jobs, extensions, and multi-surface operation.	GitHub, site	Medium	Treat IronClaw as the security parity gauntlet: breadth is not acceptable until isolation, permission, credential, and prompt-injection boundaries are proven.	M2, M3, M5, M7, M9
Claude Code	Agentic coding system across terminal, IDE, desktop, browser, and web/cloud surfaces; reads codebases, edits files, runs commands, uses MCP, skills, hooks, subagents, CI/GitHub/GitLab flows, and managed environments for cloud work.	Overview, Anthropic docs, security	High	Seraph is behind on polished coding-agent execution and multi-surface developer workflow unless task benchmarks prove otherwise.	M2, M3, M5, M7, M9
Codex	OpenAI coding agent with cloud delegated tasks and local CLI/IDE paths; reads, modifies, runs code, creates PRs, runs in task-scoped cloud containers, and supports approval modes locally.	Codex cloud, CLI help, Codex repo	High	Seraph needs coding-task execution, background delegation, sandbox, PR, and review benchmarks before any Codex-class claim.	M0, M2, M3, M5, M7
OpenHands	Open-source/cloud software-agent platform with SDK, CLI, local GUI/web GUI, Docker/process/remote sandbox modes, terminal, browser, app preview, VS Code/file editor, and enterprise/cloud coding-agent positioning.	Product, docs, GitHub, sandbox overview, runtime architecture	High	Seraph must match engineering-task workbench clarity, sandbox options, and transparent artifact/change review outside a coding-only niche.	M2, M3, M5, M7, M9
Goose	Open-source native agent with desktop, CLI, API, broad MCP extension ecosystem, provider choice, recipes, subagents, MCP Apps, Code Mode, and documented security controls.	Docs/home, GitHub, Code Mode, Apps extension	High	Goose pressures Seraph on local/native distribution, MCP breadth, recipes, subagents, and extension UX. Seraph must win with stronger governance and guardian context.	M1, M2, M3, M4, M7, M9
Aider	Focused terminal AI pair programmer with repo map, multi-file edits, git diffs/commits, broad model support, and benchmark-friendly coding workflow.	Docs, site, GitHub	High	Aider is a narrow but serious terminal coding baseline. Seraph should not claim coding superiority without diff, test, repo-map, and commit-quality benchmarks.	M0, M2, M7
Cline / Roo Code	Cline is an editor/terminal coding agent with file edits, terminal/browser actions, MCP, checkpoints, rules/skills/workflows, Memory Bank, subagents, and explicit approval framing. Roo Code is similar Cline-derived/editor-agent pressure, but public docs say Roo Code products sunset on 2026-05-15.	Cline docs, Cline overview, Roo docs, Roo GitHub	High for Cline, Medium for Roo	Seraph needs comparable IDE/editor-agent execution quality, approval ergonomics, MCP use, checkpoints, and task transparency. Roo should be treated as a changing/deprecating baseline.	M0, M2, M5, M7, M9
Devin	Managed autonomous software engineer for engineering tickets, bugs, features, tests, internal tools, terminal workflows, repository setup, and team backlog execution.	Docs	High	Devin pressures Seraph on long-running engineering-task endurance, managed task UX, repo setup, test/fix loops, and human handoff.	M0, M2, M3, M5, M7
Manus	General AI agent API for asynchronous tasks with projects, files, webhooks, skills, connectors, custom agents, task follow-ups, and task visibility controls.	API overview, task.create, website management	High	Manus pressures Seraph on general task delegation, connector/skill-controlled runs, task APIs, and consumer-facing task completion.	M2, M4, M5, M7, M9
Browserbase / Stagehand	Browser-agent infrastructure with cloud browser sessions, search/fetch, functions, model gateway, session inspector, Stagehand natural-language browser automation, Playwright/Puppeteer/Selenium support, and MCP server.	Browserbase docs, Stagehand docs, Stagehand GitHub, Browserbase MCP	High	Seraph is behind on production browser-agent infrastructure, session observability, replay, and browser automation reliability unless it integrates or matches these patterns.	M2, M3, M4, M7, M9
AutoGPT / Forge	Agent-platform lineage with self-hosted AutoGPT platform, block/workflow-style automations, Forge toolkit, Agent Protocol, benchmark culture, UI, CLI, and continuous agents.	AutoGPT GitHub, Forge protocols	Medium	Use AutoGPT/Forge primarily as M0/M1/M9 pressure for agent protocols, benchmarking, agent construction, and platform history; avoid current parity claims without narrower refresh.	M0, M1, M5, M9
CrewAI	Multi-agent framework for agents, crews, flows, tools, memory, knowledge, guardrails, observability, human-in-the-loop triggers, and enterprise automations.	Docs, introduction, agents, tools, processes	High	CrewAI pressures Seraph on mature multi-agent/workflow framework shape, enterprise automation, reusable agents, flows, and governance vocabulary.	M1, M5, M7, M9
LangGraph	Low-level orchestration framework/runtime for long-running stateful agents with durable execution, streaming, human-in-the-loop, checkpoint persistence, memory, time travel, interrupts, replay/fork, and fault tolerance.	Overview, persistence, workflows and agents, memory, GitHub	High	LangGraph sets the durable workflow primitive bar. Seraph needs checkpoint, branch, replay, fork, interrupt, and state inspection semantics that are productized for operators.	M5, M6, M7, M9

Cross-Competitor Gap Map

Milestone	Competition truth to answer first
M0	Keep this table and the claim ledger current before any world-class, best, superior, secure, private, production-ready, or ahead claim is repeated.
M1	Hermes, Goose, CrewAI, AutoGPT/Forge, LangGraph, and OpenClaw all pressure Seraph to make capability contracts, manifests, taxonomy, and contribution boundaries explicit.
M2	Claude Code, Codex, OpenHands, Aider, Cline, Devin, Hermes, Browserbase/Stagehand, and Manus set the execution-quality bar for terminal, files, code, browser, patching, tests, and artifacts.
M3	IronClaw, OpenHands, Codex cloud, Claude Code cloud, Browserbase, Hermes, and OpenClaw force Seraph to prove isolation, permissions, credential boundaries, prompt-injection defenses, and provider trust path by path.
M4	Hermes, OpenClaw, Goose, Manus, Browserbase, and Claude Code pressure Seraph on selective native/browser/messaging/API reach without trust fragmentation.
M5	LangGraph, Devin, Claude Code, Codex, OpenHands, CrewAI, Hermes, and AutoGPT/Forge pressure Seraph on durable jobs, background work, delegation, checkpointing, resume, replay, and repair.
M6	Hermes and LangGraph set bounded/contextual memory expectations; Seraph's chance to win is behavior-changing guardian memory, not larger memory inventory.
M7	OpenClaw, OpenHands, Claude Code, Devin, Browserbase, Cline, Goose, and Hermes pressure Seraph's cockpit to expose execution, approvals, artifacts, failures, spend, and recovery faster than source diving.
M8	Few competitors are explicitly guardian-first. Seraph's differentiator must be proven by salience, restraint, intervention timing, goal continuity, and feedback-conditioned capability choice.
M9	Goose, Hermes, CrewAI, LangGraph, AutoGPT/Forge, Cline, OpenClaw, and Browserbase pressure Seraph's ecosystem to scale through governed packages, managed connectors, review, compatibility, and trust levels.

Claim Rules

• This file may support Behind, At par, Unknown, Partially backed, or aspirational strategy language.
• It does not by itself support Ahead, Best, World-class, Secure, Private, Production-ready, or Superior claims.
• Any such claim must also pass the status and allowed-wording rules in 19. Strategy Claim Ledger.
• If a competitor row depends on non-official or stale evidence, keep confidence below High and phrase the gap as a watch item.