company harness
paperclip
turns many agents into a governed operating system: goals, org chart, budgets, tickets, audit, heartbeats.
orchestration proof
49.5k stars
one-screen read
three harnesses, three different lessons
the useful move is not to pick a winner. it is to see what each one solved that the others did not.
augmentation harness
oh-my-codex
does not replace codex. wraps it with skills, hooks, roles, team runtime, hud, and durable local state.
workflow proof
18.5k stars
personal runtime
hermes-agent
pushes hard on long-lived personal agenthood: memory, skills, learning loop, messaging surfaces, scheduling.
continuity proof
33.2k stars
cards
what each harness teaches
each card is structured as thesis -> mechanism -> what to steal -> what breaks on transfer.
paperclip
the company as interface
paperclip’s core move is not “more agents.” it is turning agents into a managed company with goals, reporting lines, budgets, tickets, and governance.
why its reputation is real
the repo has breakout traction, the landing page is coherent, and your own repo already uses it as an execution and audit layer for investor write/eval loops.
- mechanism: bring-your-own agents, org chart, heartbeats, ticket threads, immutable audit, budget throttling, goal ancestry.
- what to steal: ticketed execution, audit trail, approval boundaries, scheduled wakeups, explicit governance.
- what not to import blindly: “zero-human company” framing and org-chart maximalism. secondme is principal-centric, not business-sim theater.
- why it matters for secondme: it proves the control-plane can be a product surface, not just hidden plumbing.
oh-my-codex
the agent is not alone
omx assumes codex is already useful, then asks: what workflow layer, team runtime, and project-local memory make the same engine much stronger in practice?
why its reputation is real
fast traction, clear product thesis, strong docs discipline, and a very legible framing: codex stays the engine, omx upgrades the runtime around it.
- mechanism: hooks, roles, skills, `$deep-interview`, `$ralplan`, `$team`, `$ralph`, and durable `.omx/` state for plans/logs/memory.
- what to steal: workflow-first augmentation, codified planning rituals, team mode, project-local persistent runtime layer.
- what not to import blindly: coding-task bias and feature-surface inflation. secondme cannot become “codex, but with more commands.”
- why it matters for secondme: it proves there is leverage in wrapping a strong execution engine with better ritual, memory, and coordination.
hermes-agent
the personal runtime that accumulates self
hermes-agent is the strongest current push toward a persistent personal agent that lives across sessions, surfaces, and memory layers instead of dying at every prompt boundary.
why its reputation is real
nous has brand gravity, the repo has strong pull, and the product thesis is unusually aligned with continuity: memory, skills, learning loop, messaging, automations, subagents.
- mechanism: persistent memory, user model, self-created skills, gateway across telegram/discord/slack/etc, cron, MCP, parallel subagents, cloud-resident runtime.
- what to steal: cross-surface continuity, memory as active substrate, messaging-native operation, long-lived agent home outside the laptop.
- what not to import blindly: self-improvement mystique. if learning is not inspectable, it becomes narrative instead of trust.
- why it matters for secondme: it is the closest of the three to the “chief-of-staff that remembers you over time” axis.
compare
where each one sits in the stack
the key is to stop comparing them as if they solve the same job. they do not.
paperclip wins at
control plane
governance, approval, budget, task routing, agent fleet management, auditability.
company brainstem
oh-my-codex wins at
workflow wrapper
taking an already-good agent engine and making the day-to-day operating ritual stronger.
runtime augmentation
hermes-agent wins at
continuity runtime
identity, memory, messaging, scheduling, and “lives with you” semantics.
personal continuity
paperclip weakness
too company-shaped
strong on governance, weaker as a model of one principal’s identity, mood, rhythm, and cognitive light cone.
oh-my-codex weakness
too coding-shaped
great for implementation loops; not automatically a life / chief-of-staff substrate.
hermes-agent weakness
trust can go fuzzy
the more it says “self-improving,” the higher the burden for inspectability, policy, and boundedness.
transfer rules
what secondme should actually take
the synthesis is not a blend. it is a stack.
1
from paperclip: make the control plane visible
tickets, approvals, traces, wakeups, and governance should not be hidden implementation detail. they are part of trust.
2
from oh-my-codex: treat workflow as a first-class harness
the engine can stay the engine. leverage comes from the layer that structures clarification, planning, execution, and coordination.
3
from hermes-agent: continuity must outlive the session
memory, messaging surfaces, skills, and user modeling should converge into one long-lived personal runtime.
4
the real stack for secondme
paperclip spine for governance, omx-style wrapper for workflow quality, hermes-like continuity for identity and memory. that composite is much closer to the product than any one source alone.
one step
what to prototype next
if this page is useful, the next move is not more reading. it is one synthetic spike.
spike a
paperclip-style ticket + trace loop
take one secondme workflow and force it through issue documents, wakeups, and explicit approval checkpoints.
spike b
omx-style ritual wrapper
turn one workflow into a fixed ritual: clarify, approve plan, execute, synthesize, persist state.
spike c
hermes-style continuity layer
make the same principal visible across chat, telegram, memory, and scheduled briefings without re-explaining reality each time.