control plane
paperclip
visible governance for agent work: goals, org charts, budgets, wakeups, tickets, audit.
transfer fit high
50.2k stars
breakout
core dashboard
the 5 core harnesses worth tracking now
this is the working shortlist. it mixes control plane, wrapper, continuity, memory, and cognition on purpose because those are the layers our product must eventually fuse.
workflow wrapper
oh-my-codex
shows that a strong engine becomes much better when wrapped in ritual, hooks, roles, teams, and local durable state.
transfer fit high
19.3k stars
docs-led
continuity runtime
hermes-agent
the closest current runtime to “one agent that keeps living after the session dies.”
continuity high
39.3k stars
migration pull
memory substrate
supermemory
the memory layer that feels closest to our context-engineering direction: hydration, shared memory, plugins, mcp, and cross-runtime recall.
memory pressure
21.6k stars
ecosystem
cognitive harness
arcgentica
the strongest current proof trigger for `harness > model`: decomposition, orchestrator-subagent flow, and persistent runtime on hard reasoning tasks.
cognition pressure
36% arc result
repo signal weak
signals
what each one contributes to the stack
the point is not to choose a winner. the point is to see which missing mechanic each source exposes and which loop it strengthens.
paperclip
make the control plane visible
strongest proof that approvals, wakeups, issue threads, and governance should be product surface, not invisible plumbing.
oh-my-codex
the wrapper can beat the raw engine
best proof that workflow rituals, hooks, and team runtime can compound a strong base agent without replacing it.
hermes-agent
continuity must outlive the laptop
best current pressure source for memory, messaging, scheduling, and long-lived runtime semantics.
supermemory
memory should hydrate, not just store
best current pressure source for context packaging, shared recall across runtimes, and memory as living substrate instead of dead archive.
arcgentica
cognition is also a harness problem
best current pressure source for orchestrator-subagent reasoning, decomposition, persistent runtime, and compression over raw context.
research cards
what each live harness says right now
every card uses the same contract: thesis, mechanism, community read, transfer move, and rejection rule.
paperclip
governed orchestration
turns a fleet of agents into a managed operating system instead of a bag of prompts.
community read
breakout github signal is real. in this repo it is already more than inspiration because we use it for eval and writing loops.
- mechanism: org charts, budgets, goals, tickets, wakeups, traceable execution.
- borrow: visible approvals, issue-centric execution, explicit audit.
- reject: zero-human-company metaphysics and company-sim maximalism.
trust fit high
transfer fit high
maintenance medium
oh-my-codex
workflow-first augmentation
assumes the engine already works, then upgrades how planning, coordination, and memory behave around it.
community read
the signal is mostly github and docs quality, not broad reddit chatter. that is fine. it still proves the runtime wrapper thesis.
- mechanism: hooks, roles, skills, team mode, persistent local runtime layer.
- borrow: codified rituals, agent teams, local state, workflow hardening.
- reject: coding-task bias and feature-surface inflation.
transfer fit high
trust fit medium
maintenance low
hermes-agent
personal continuity runtime
closest current pressure source for a persistent personal agent that spans memory, messaging, schedules, and subagents.
community read
strong repo pull plus visible migration chatter from
openclaw users who want lower maintenance or better continuity.
- mechanism: persistent memory, user model, gateway surfaces, schedules, skills, parallel subagents.
- borrow: continuity outside sessions, messaging-native runtime, living agent home.
- reject: self-improvement mystique without inspectable learning.
continuity high
transfer fit medium
trust risk medium
supermemory
memory as hydration layer
the closest current memory stack to our framing: not only recall, but a distributed context layer across tools and agent runtimes.
community read
good repo traction plus an integration ecosystem around Claude, MCP, and OpenClaw. useful because it is not only a memory api, but a context-distribution posture.
- mechanism: memory api, developer plugins, claude integration, mcp bridge, shared recall across runtimes.
- borrow: hydration, memory portability, context packaging, cross-runtime memory layer.
- reject: the idea that memory substrate alone solves judgment, trust, or doctrine.
memory fit high
trust fit medium
maintenance low
arcgentica
cognition through orchestration
important because it proves the model is not the full unit of intelligence. the harness can massively change reasoning quality on hard tasks.
community read
weak github traction on the repo itself, but strong conceptual weight because of the benchmark result and the architectural lesson attached to it.
- mechanism: orchestrator-subagent architecture, compressed briefs, persistent runtime objects, code-mode reasoning.
- borrow: decomposition, adaptive specialist routing, runtime-as-context, problem-solving harness logic.
- reject: assuming benchmark transfer without local workflow replication.
cognition fit high
loop fit medium
transfer proof low
forgotten but relevant
the sources that should not be mixed into the same bucket
these matter a lot, but they are principle layers or pressure layers, not the same class as the core harnesses.
boundary source
daniel miessler
not a harness. category language, scaffolding-over-model doctrine, personal ai infrastructure framing.
source layer
doctrine high
boundary source
symbolica
keep this as the principle layer behind arcgentica: runtime-as-context, typed tools, trace graphs, persistent objects, benchmark-driven proof.
architecture pressure
runtime state
anti-pattern
openclaw
still huge in community pull, but useful here mainly as maintenance-tax and migration-pressure evidence, not as a clean inspiration source.
maintenance tax
352.5k stars
architecture contrast
openclaw vs hermes vs secondme
this should become a permanent architecture section, because it explains both the market wedge and the trust difference.
| system | what it proved | where it breaks | what secondme changes |
|---|---|---|---|
| openclaw | showed the surface potential of a personal agent that can act across contexts and tools | maintenance tax, weak context management, goal forgetting, low trust in long loops | we shift from infra burden to bounded leverage: inspectable state, better context management, explicit approvals, lower operator pain |
| hermes-agent | showed a stronger continuity runtime and a more ambitious self-improvement story | self-improvement may still be too fuzzy; trust burden rises if learning is not inspectable enough | we care about inspectable memory governance and explicit learning loops tied to user-reported failures and accepted edits |
| secondme | target state: chief-of-staff harness for context management, bounded action, and compounding principal-specific continuity | not yet proven locally across a frozen workflow | must combine control plane, continuity, hydration, cognition, and inspectable self-improvement without inheriting the maintenance tax |
dossier plan
how to run the deep research per harness
each dossier should be comparable. otherwise we will get five different essays and no actual transfer logic.
mandatory dossier contract
same shape for every harness
- 1. what it is and who it is for.
- 2. why the community cares, with a distinction between repo traction and real user chatter.
- 3. the architectural primitives: runtime, memory, permissions, traces, orchestration.
- 4. the self-improvement loop: does it learn, how, and can we inspect it.
- 5. what secondme should steal, what it should reject, and what local spike tests transfer.
- 6. what evidence would promote it into doctrine instead of keeping it as pressure source only.
recommended order
run these first
- now: `paperclip`, `hermes-agent`, `arcgentica`.
- next: `oh-my-codex`, `supermemory`.
- watch-only: `openclaw`, `symbolica`, `miessler`.
why this order
first resolve control plane, continuity, and cognition. then refine wrapper ritual and memory hydration.
white paper use
what this research is already saying
the point is not only to compare tools. the point is to extract the architecture claims strong enough to carry into the white paper.
claim 1
the control plane is part of trust
paperclip is the strongest proof that tickets, approvals, wakeups, and traces should be visible product surface, not hidden backend machinery.
claim 2
continuity beats session sharpness
hermes-agent and supermemory both point in the same direction: the user experiences value when the system starts from living context instead of cold prompts.
claim 3
cognition is also a harness problem
arcgentica is the strongest current proof trigger that better decomposition, specialist routing, and runtime state can change reasoning quality materially.
claim 4
self-improvement needs inspection
hermes-agent is useful precisely because it pressures this question: does the system improve only by writing skills, or by safely changing how it works over time?
claim 5
memory is not just retrieval
supermemory is most useful where it distinguishes temporal memory and profile hydration from plain document search or generic rag.
claim 6
maintenance tax kills the dream
openclaw still matters because it shows the gap between perceived potential and lived operator burden. that is a wedge, not a side note.
comparison map
what secondme should actually take
the product should not clone any one source. it should take one hard mechanic from each winning layer.
| source | strongest mechanic | what we borrow | what we reject |
|---|---|---|---|
| paperclip control plane | ticketed, audited, wakeup-driven execution | visible approvals, traces, issue documents, governance | company-sim maximalism and zero-human rhetoric |
| oh-my-codex wrapper layer | workflow ritual around a strong base engine | hooks, roles, skills, plan discipline, team runtime | coding-shaped surface area as product identity |
| hermes-agent continuity layer | one runtime that survives across surfaces | persistent state, messaging surfaces, schedules, skills | fuzzy self-improvement story without inspectability |
| supermemory memory substrate | portable shared memory and context hydration across runtimes | hydration, portability, context packaging, plugin ecosystem | equating memory layer with full judgment layer |
| arcgentica cognition layer | decomposition and persistent-runtime problem solving | compressed specialist briefs, orchestration, runtime-as-context | assuming benchmark wins transfer without workflow replication |
repo metrics checked on 2026-04-09. community reads are a mix of github traction, official docs/sites, local secondme notes, and sampled public threads around `openclaw -> hermes`, `supermemory`, and `arcgentica` benchmark discourse.