Personal AgentsSources & methodology.
What is documented, what is inferred, and what is still unknown.
Quality ratings, not feature counts
Grades synthesize cited firsthand outcomes and documented trade-offs. They are our estimates, not the source authors’ scores. No cross-study average is computed. Different tasks, dates, models, and configurations limit comparison. A 1–2 point gap may not be meaningful. Missing evidence is unscored, never zero.
The charts are our editorial research synthesis. We have not personally run an identical task suite on all eight products. We do not pretend scores from different authors, dates, or task sets form one controlled benchmark.
Vector length is the only data quantity. Angular direction identifies the agent; perspective depth and vector height do not introduce extra scores.
Moderate means corroborating or detailed firsthand evidence. Limited means narrow tests, a single opinion, configuration dependence, or design inference. No grade has high-confidence universal validity.
Control and ownership: these slides use explicitly labeled design-and-operations assessments. They are not safety audit results. Self-hosting does not automatically make an agent safe or private.
Rubric: 0–2: Poor / major failures · 3–4: Limited / substantial friction · 5–6: Mixed / repeated correction · 7–8: Good / useful with review · 9–10: Strong in the reviewed evidence
Performance evidence
Diana’s Claude Duet business-task test ↗
Firsthand task review · 2026-10-28
Two days; different briefs from earlier agent tests. Nine scored tasks; the failed audio task is excluded from the published average.
Diana’s Dots business-task test ↗
Firsthand task review · 2026-10
Editorial task scoring, not a controlled cross-product benchmark. Spreadsheet self-checks did not guarantee a working output.
Diana’s Muse business-task test ↗
Firsthand task review · 2026-10
Some tasks were stopped for access boundaries or left inconclusive; do not treat omitted tasks as passes.
Diana’s Gemini Spark task test ↗
Firsthand task review · 2026
Google-centric business tasks; earlier Spark experience, not a test of every current Gemini task feature.
Independent hands-on scorecard · 2026-11-07
Eight of fifteen categories tested. Limited personal-account testing, not a representative user sample.
Independent hands-on scorecard · 2026-10-31
Eight of fifteen categories tested; not every feature or workflow was evaluated.
Assistant Benchmark: Instinct ↗
Independent hands-on scorecard · 2026-11-07
Twelve of fifteen categories tested. Results are specific to the reviewer’s tasks and accounts.
Assistant Benchmark: Grok Bot ↗
Independent hands-on scorecard · 2026-11-07
Eleven of fifteen categories tested. Other testers report different email and workflow experiences.
PersonalAgentEval: real-product task evaluation ↗
Published evaluation · 2026-10
152 tasks with a simulated user and test accounts; small memory subgroups. Duet, Hermes, and OpenClaw are not in this cohort.
PersonalAgentEval: task design and limitations ↗
Evaluation methodology · 2026-10-14
Controlled synthetic personas are not equivalent to a live user’s whole digital life.
Reece Rogers: six-agent, one-day test ↗
Firsthand comparative test · 2026-10-27
One tester, five tasks, one day. Early-release results, not a longitudinal reliability or security assessment.
Peter Yang: four-agent firsthand comparison ↗
Firsthand comparative opinion · 2026-10-21
Qualitative opinion and demonstration, not controlled measurement. Related video has a sponsor.
Julian Goldie: Hermes vs OpenClaw ↗
Firsthand comparative opinion · 2026-10-20
Commercial creator’s subjective scores, not a formal benchmark; lower confidence than independent repeated tests.
DreamHost: Docker setup and migration experiment ↗
Published operational experiment · 2026-10-22
One arm64 VM; no model connected. Measures setup, resources, and migration—not answer quality or task success. Hosting vendor.
Anthropic: Duet limits and controls ↗
Vendor documentation · 2026-11-07
Product design claims, not independent performance or safety testing.
Meta: Muse architecture and safeguards ↗
Vendor documentation · 2026-11-07
Vendor-described design; not proof of real-world safety.
OpenAI: Dots controls and usage ↗
Vendor documentation · 2026-11-07
Product design and access constraints, not a controlled performance result.
Vendor documentation · 2026-11-07
Documented workflow and runtime design, not measured reliability.
Google: tasks and workflows in Gemini ↗
Vendor documentation · 2026-11-07
Current availability and design. Earlier reviews may refer to the separate Spark surface.
Vendor documentation · 2026-11-07
Runtime options and design; results depend on the selected model, tools, and configuration.
OpenClaw: gateway security and control ↗
Vendor documentation · 2026-11-07
Configuration guidance, not a security audit or an outcome benchmark.
Instinct: product and access ↗
Vendor documentation · 2026-11-07
Limited public technical detail; no ownership or export performance test found.
Initiative — grades, rationale, and evidence
Useful unsolicited help, routine follow-through, and the amount of prompting still needed.
Claude Duet · 9/10 · Moderate evidence · Outcome synthesis
Strong proactive judgment; longer-term consistency remains early evidence.
Peter Yang: four-agent firsthand comparison · Reece Rogers: six-agent, one-day test
Muse · 7/10 · Moderate evidence · Outcome synthesis
Routines work well; spontaneous cross-app problem detection is less consistent.
Diana’s Muse business-task test · Assistant Benchmark: Muse · Reece Rogers: six-agent, one-day test
Dots · 9/10 · Moderate evidence · Outcome synthesis
Strong observed conflict detection and recurring-job follow-through.
Diana’s Dots business-task test · Reece Rogers: six-agent, one-day test
Instinct · 7/10 · Moderate evidence · Outcome synthesis
Useful routines, but not every important issue is surfaced without explicit setup.
Assistant Benchmark: Instinct · Reece Rogers: six-agent, one-day test
Grok Bot · 7/10 · Moderate evidence · Outcome synthesis
Repeatable routines are strong; proactive judgment still needs direction.
Assistant Benchmark: Grok Bot · Reece Rogers: six-agent, one-day test
Gemini Spark · 5/10 · Limited evidence · Outcome synthesis
Good requested briefs; weaker evidence for finding important work unaided.
Diana’s Gemini Spark task test · Reece Rogers: six-agent, one-day test
Hermes Agent · 8/10 · Limited evidence · Outcome synthesis
Strong scheduled automation; score reflects configured routines, not default initiative.
Julian Goldie: Hermes vs OpenClaw · DreamHost: Docker setup and migration experiment
OpenClaw · 7/10 · Limited evidence · Outcome synthesis
Useful always-on workflows, with more configuration and maintenance responsibility.
Julian Goldie: Hermes vs OpenClaw · OpenClaw: gateway security and control
Memory & adaptation — grades, rationale, and evidence
Useful recall, preference application, and adaptation—not simply having a memory feature.
Claude Duet · 8/10 · Limited evidence · Outcome synthesis
Prior context improved outputs, but the review covers only a short period.
Diana’s Claude Duet business-task test
Muse · 8/10 · Moderate evidence · Outcome synthesis
Strong narrow recall results; personalization is less consistent across business outputs.
PersonalAgentEval: real-product task evaluation · Diana’s Muse business-task test
Dots · 8/10 · Moderate evidence · Outcome synthesis
Strong recall overall; context updates and follow-through still show gaps.
PersonalAgentEval: real-product task evaluation · Assistant Benchmark: Dots · Diana’s Dots business-task test
Instinct · 9/10 · Moderate evidence · Outcome synthesis
Strong preference and context retention in the published test cohort.
PersonalAgentEval: real-product task evaluation · Assistant Benchmark: Instinct
Grok Bot · 8/10 · Limited evidence · Outcome synthesis
Excellent small-sample recall; limited evidence for broader long-term learning.
PersonalAgentEval: real-product task evaluation
Gemini Spark · 6/10 · Limited evidence · Outcome synthesis
Memory performance was weaker than peers in the evaluated subset.
PersonalAgentEval: real-product task evaluation
Hermes Agent · 9/10 · Limited evidence · Outcome synthesis
Experience-to-skill learning is a firsthand reviewer’s principal advantage.
Julian Goldie: Hermes vs OpenClaw
OpenClaw · 7/10 · Limited evidence · Outcome synthesis
Useful persistent context, but less convincing cumulative learning in that comparison.
App workflows — grades, rationale, and evidence
Completing multi-app work cleanly, handling account context, and returning usable results.
Claude Duet · 8/10 · Moderate evidence · Outcome synthesis
Strong connected-account judgment; account switching adds friction.
Diana’s Claude Duet business-task test
Muse · 7/10 · Moderate evidence · Outcome synthesis
Excellent everyday integrations in one test; business file and login boundaries still interfere.
Assistant Benchmark: Muse · Diana’s Muse business-task test
Dots · 8/10 · Moderate evidence · Outcome synthesis
Broadly useful connected workflows; email and spreadsheet correction remains necessary.
Diana’s Dots business-task test · Assistant Benchmark: Dots
Instinct · 8/10 · Moderate evidence · Outcome synthesis
Strong real-life task completion; results still vary by errand and account.
PersonalAgentEval: real-product task evaluation · Assistant Benchmark: Instinct
Grok Bot · 8/10 · Moderate evidence · Outcome synthesis
Capable cross-app work, but testers disagree on email effectiveness.
Assistant Benchmark: Grok Bot · Peter Yang: four-agent firsthand comparison
Gemini Spark · 8/10 · Moderate evidence · Outcome synthesis
Strongest when the workflow stays inside Google’s native apps.
Diana’s Gemini Spark task test
Hermes Agent · 8/10 · Limited evidence · Outcome synthesis
Flexible workflows once configured; the positive assessment is from a commercial reviewer.
Julian Goldie: Hermes vs OpenClaw
OpenClaw · 8/10 · Limited evidence · Outcome synthesis
Strong integration ecosystem, with more setup burden than managed products.
Computer use — grades, rationale, and evidence
Observed browser and computer-task execution, not possession of a virtual machine.
Claude Duet · Unscored · Unscored evidence · Insufficient evidence
No sufficiently comparable browser execution test found. Product limits alone do not establish quality.
Anthropic: Duet limits and controls · Diana’s Claude Duet business-task test
Muse · 8/10 · Moderate evidence · Outcome synthesis
Strong online-action score; credentials and handoffs can interrupt completion.
Assistant Benchmark: Muse · Diana’s Muse business-task test
Dots · 8/10 · Moderate evidence · Outcome synthesis
Strong execution, with occasional workflow and artifact failures.
Assistant Benchmark: Dots · Diana’s Dots business-task test
Instinct · 7/10 · Moderate evidence · Outcome synthesis
Useful online execution, but not reliably hands-off in every errand.
Assistant Benchmark: Instinct · PersonalAgentEval: real-product task evaluation
Grok Bot · 8/10 · Limited evidence · Outcome synthesis
Good browser-based work; some tasks still require intervention.
Gemini Spark · Unscored · Unscored evidence · Insufficient evidence
Available browsing features were not enough to establish comparable execution quality.
Google: tasks and workflows in Gemini
Hermes Agent · Unscored · Unscored evidence · Insufficient evidence
The operational test used no model; it cannot support a browser-quality score.
DreamHost: Docker setup and migration experiment · Nous Research: Hermes Agent
OpenClaw · Unscored · Unscored evidence · Insufficient evidence
Configuration documentation is not an end-to-end browser performance test.
OpenClaw: gateway security and control · DreamHost: Docker setup and migration experiment
Communication — grades, rationale, and evidence
Clarity, interaction flow, and practical usability of the surfaces reviewers actually tried.
Claude Duet · 9/10 · Moderate evidence · Outcome synthesis
Clear, polished task handling stood out to multiple firsthand reviewers.
Peter Yang: four-agent firsthand comparison · Reece Rogers: six-agent, one-day test
Muse · 9/10 · Limited evidence · Outcome synthesis
Consumer interaction and business calling received strong firsthand praise.
Peter Yang: four-agent firsthand comparison
Dots · 7/10 · Limited evidence · Outcome synthesis
Useful familiar context, but slow voice connections reduce fluidity.
Peter Yang: four-agent firsthand comparison
Instinct · 6/10 · Limited evidence · Outcome synthesis
Text-first simplicity suits some users; limited visibility frustrated another tester.
Reece Rogers: six-agent, one-day test · Assistant Benchmark: Instinct
Grok Bot · 9/10 · Limited evidence · Outcome synthesis
Voice and the specialist-bot interaction model earned strong firsthand praise.
Peter Yang: four-agent firsthand comparison
Gemini Spark · 6/10 · Limited evidence · Outcome synthesis
Communication and context handling were mixed in the one-day comparison.
Reece Rogers: six-agent, one-day test
Hermes Agent · 8/10 · Limited evidence · Outcome synthesis
Positive messaging experience in a creator’s comparison; model and channel setup matter.
Julian Goldie: Hermes vs OpenClaw
OpenClaw · 8/10 · Limited evidence · Outcome synthesis
A straightforward chat assistant in the same comparison; operator setup still matters.
Creation — grades, rationale, and evidence
Correctness, polish, and usability of documents, websites, spreadsheets, and creative outputs.
Claude Duet · 9/10 · Moderate evidence · Outcome synthesis
Strong finished business work; audio transcription failed and previews had limits.
Diana’s Claude Duet business-task test
Muse · 6/10 · Moderate evidence · Outcome synthesis
Research and briefs were useful; webpage and spreadsheet delivery were uneven.
Diana’s Muse business-task test
Dots · 9/10 · Moderate evidence · Outcome synthesis
Strong research and webpage outputs; spreadsheet correctness was weaker.
Diana’s Dots business-task test
Instinct · Unscored · Unscored evidence · Insufficient evidence
Errand success does not establish document, code, or media-production quality.
Grok Bot · 8/10 · Limited evidence · Outcome synthesis
Good creative work and Pages feedback, with less detailed artifact testing.
Assistant Benchmark: Grok Bot · Peter Yang: four-agent firsthand comparison
Gemini Spark · 8/10 · Moderate evidence · Outcome synthesis
Strong briefs and plans; webpage preview friction reduces end-to-end usefulness.
Diana’s Gemini Spark task test
Hermes Agent · Unscored · Unscored evidence · Insufficient evidence
No sufficiently comparable finished-artifact evaluation found.
OpenClaw · Unscored · Unscored evidence · Insufficient evidence
Showcases demonstrate possibilities, not a dependable artifact-quality score.
Delegation — grades, rationale, and evidence
Successful task handoffs, parallel coordination, and the effort of directing specialists.
Claude Duet · Unscored · Unscored evidence · Insufficient evidence
Topic channels are not a tested specialist-agent team.
Anthropic: Duet limits and controls
Muse · Unscored · Unscored evidence · Insufficient evidence
Subagents are documented, but independent coordination quality is not established here.
Meta: Muse architecture and safeguards
Dots · 8/10 · Limited evidence · Outcome synthesis
A useful orchestrator of existing work threads, with early-product rough edges.
Peter Yang: four-agent firsthand comparison
Instinct · Unscored · Unscored evidence · Insufficient evidence
No comparable specialist coordination test found.
Grok Bot · 9/10 · Limited evidence · Outcome synthesis
The specialist-team model was especially clear to a comparative tester.
Peter Yang: four-agent firsthand comparison · Assistant Benchmark: Grok Bot
Gemini Spark · Unscored · Unscored evidence · Insufficient evidence
Tasks and skills alone do not establish multi-agent coordination quality.
Google: tasks and workflows in Gemini
Hermes Agent · Unscored · Unscored evidence · Insufficient evidence
Subagent support is documented; controlled coordination outcomes were not found.
OpenClaw · Unscored · Unscored evidence · Insufficient evidence
Configurable routing is not proof of reliable multi-agent handoffs.
Control & transparency — grades, rationale, and evidence
Understandable permissions, review surfaces, and practical visibility—not resistance to every security attack.
Claude Duet · 8/10 · Limited evidence · Design inference
Clear review-oriented design; the cited task test did not exercise every approval flow.
Anthropic: Duet limits and controls · Diana’s Claude Duet business-task test
Muse · 7/10 · Limited evidence · Design inference
Useful control surfaces, with consequential account access still a significant trade-off.
Assistant Benchmark: Muse · Meta: Muse architecture and safeguards
Dots · 8/10 · Limited evidence · Design inference
Review and pause controls are helpful; correctness still needs human checking.
Assistant Benchmark: Dots · OpenAI: Dots controls and usage
Instinct · 6/10 · Limited evidence · Design inference
Permission scores were positive, but a firsthand tester found activity visibility weak.
Assistant Benchmark: Instinct · Reece Rogers: six-agent, one-day test
Grok Bot · 8/10 · Limited evidence · Design inference
Inspectability is useful; shared files and sign-ins need careful boundaries.
Assistant Benchmark: Grok Bot · Grok Bot: computer and apps
Gemini Spark · 7/10 · Limited evidence · Design inference
Task review and connected-account controls are clear; independent safety evidence is limited.
Google: tasks and workflows in Gemini
Hermes Agent · 7/10 · Limited evidence · Design inference
Local control can be strong, but configuration and model-provider decisions stay with you.
Nous Research: Hermes Agent · DreamHost: Docker setup and migration experiment
OpenClaw · 7/10 · Limited evidence · Design inference
Detailed operator controls; secure operation requires correct policy and gateway setup.
Ownership & portability — grades, rationale, and evidence
Control over runtime, model choice, and moving useful state—not a claim that self-hosting is automatically private.
Claude Duet · 3/10 · Limited evidence · Design inference
Convenient managed service; runtime and model choice remain vendor-controlled.
Anthropic: Duet limits and controls
Muse · 3/10 · Limited evidence · Design inference
A dedicated VM still belongs to a managed product ecosystem.
Meta: Muse architecture and safeguards
Dots · 3/10 · Limited evidence · Design inference
Your work is portable in parts, but the agent runtime is service-managed.
OpenAI: Dots controls and usage
Instinct · Unscored · Unscored evidence · Insufficient evidence
Public export and runtime information is too limited for a fair portability grade.
Grok Bot · 4/10 · Limited evidence · Design inference
A flexible workspace, but no self-hosted equivalent to the managed bot runtime.
Gemini Spark · 3/10 · Limited evidence · Design inference
Useful Google-native outputs, with a service-controlled runtime and account dependencies.
Google: tasks and workflows in Gemini
Hermes Agent · 9/10 · Moderate evidence · Design inference
Strong provider/runtime choice; tested migration moved useful state but needed manual cleanup.
DreamHost: Docker setup and migration experiment · Nous Research: Hermes Agent
OpenClaw · 9/10 · Moderate evidence · Design inference
Strong runtime ownership; migration worked with operational caveats.
DreamHost: Docker setup and migration experiment · OpenClaw: gateway security and control
Setup & everyday ease — grades, rationale, and evidence
Onboarding friction and practical operating burden, rather than a tally of supported platforms.
Claude Duet · 8/10 · Moderate evidence · Outcome synthesis
Polished onboarding with existing Claude context; waitlist and plan access remain constraints.
Peter Yang: four-agent firsthand comparison · Anthropic: Duet limits and controls
Muse · 9/10 · Moderate evidence · Outcome synthesis
The easiest consumer setup in the six-agent test.
Reece Rogers: six-agent, one-day test · Peter Yang: four-agent firsthand comparison
Dots · 7/10 · Limited evidence · Outcome synthesis
Familiar ChatGPT environment; remaining bugs and plan constraints add friction.
Peter Yang: four-agent firsthand comparison · OpenAI: Dots controls and usage
Instinct · 8/10 · Limited evidence · Outcome synthesis
Low-friction texting model, although not everyone likes the sparse interface.
Assistant Benchmark: Instinct · Reece Rogers: six-agent, one-day test
Grok Bot · 8/10 · Limited evidence · Outcome synthesis
An intuitive specialist-bot model, with usage and account setup to manage.
Peter Yang: four-agent firsthand comparison · Assistant Benchmark: Grok Bot
Gemini Spark · 7/10 · Moderate evidence · Outcome synthesis
Smoothest for existing Google workflows; account requirements narrow the fit.
Diana’s Gemini Spark task test · Google: tasks and workflows in Gemini
Hermes Agent · 7/10 · Moderate evidence · Outcome synthesis
Desktop improvements reduce friction; hosting, models, and permissions still require care.
Julian Goldie: Hermes vs OpenClaw · DreamHost: Docker setup and migration experiment
OpenClaw · 6/10 · Moderate evidence · Outcome synthesis
Capable but operationally heavier; configuration and migration require hands-on work.
DreamHost: Docker setup and migration experiment · Julian Goldie: Hermes vs OpenClaw