AGENT SKILL

agent-operated-software

Use when designing, building, operating, or diagnosing an ongoing application whose live backend or control loop includes OpenRig agents, including applications with a Markdown, YAML, or JSON agent control plane or a thin surface over specialist agent roles.

License 许可
Apache-2.0
最近更新
2026-11-12
Stars
★ 16
Forks
4

01 前言 Front Matter

name
agent-operated-software
metadata
{ "openrig": { "stage": "provisional", "audience": "any agent building or operating an agent-backed application", "sibling_skills": [ "agent-operated-workflows", "forming-an-openrig-mental-model", "mission-slice-sop", "openrig-user" ] } }
description
Use when designing, building, operating, or diagnosing an ongoing application whose live backend or control loop includes OpenRig agents, including applications with a Markdown, YAML, or JSON agent control plane or a thin surface over specialist agent roles.

03 SKILL.md 正文

Agent-Operated Software

Agent-Operated Software is an ongoing application whose functioning runtime includes an OpenRig agent or rig in its live backend or control loop. The application surface can be thin: structured state and specialist roles do the work, while the UI makes that work legible and steerable.

Start with the shared taxonomy

  • AI-enabled software: application code owns the control loop and calls a model as one capability.
  • Agent-Operated Workflow: an agent owns one bounded procedure with a start and stop condition; see agent-operated-workflows.
  • Agent-Operated Software: agents participate in an ongoing application's functioning runtime.

An application may compose many Agent-Operated Workflows. A bounded workflow does not by itself make its surrounding product Agent-Operated Software, and a model call is neither category unless an agent owns part of the control loop.

Agents are part of the backend

In this architecture, behavior is not located only in functions. It is distributed across agents, skills, instructions, bootstrap files, schemas, Markdown, YAML, JSON, folders, and deterministic tools. A surprising result may therefore be a code defect or a coherence gap in this control plane. Trace the path that actually produced the behavior before choosing which layer to repair.

This is not permission to tolerate defects in OpenRig core or another rock-solid substrate. There, reproduce the bug, fix the code, and hold the full gate. The faster coherence-first posture belongs to recoverable agent-operated application layers whose state and behavior can be inspected and repaired cheaply.

Markdown is the control plane

Markdown is maximally useful to an agent while remaining legible enough for a human to steer. Put intent, work state, evidence, and decisions into stable addressed artifacts rather than private chat or an opaque custom database. Keep schemas and conventions aligned with the running behavior: the agent population acts on those files as executable context.

Use progressive disclosure. A skill's name and description are the hot trigger; its body is cold procedure. Descriptions say when to load, not how to work. The body should route the reader to exact sources instead of copying them into another doctrine fork.

Scripts can act as prompts. A good script returns the context or exact effect an agent needs; when it cannot, its failure teaches what it observed, why it stopped, and safe next actions. Keep deterministic mechanics in tools and judgment in agents.

Make artifacts self-certifying

When a tool produces an artifact, do not make every consumer invent a completeness test:

  1. Write to a partial path, never the final path.
  2. Verify the property that can actually be wrong, such as duration, streams, resolution, or schema.
  3. Flush and atomically rename the partial artifact to its final path.
  4. Write the manifest, then a .done sentinel last.

Consumers of an artifact fail closed on the shared proof: don't read a file before its sentinel exists, because it may be half-written. A filename or exit code is not content verification. This rule is for artifact integrity only. For delivery, routing and checks, uncertainty fails open: deliver, and report what couldn't be checked. The hard lines are the exception: when a change can't tell whether it would expose a credential, delete someone's data, spend money or publish, it stops and asks.

Build with an agent SDLC

Select the rigor required by the outcome and the project's policy. Scope relationships, attributed evidence-backed item judgment, and genuinely distinct higher judgment have separate owners. One item judgment derives its upstream readiness; agents do not synchronize parent checkboxes, queue tags, or status prose. Capture alone is not acceptance, and readiness does not publish or advance a workflow. Use openrig-operating-model and rig proof --help for the supported judgment, correction, and read path. Keep historical locks and evidence addressable. The human supplies steering, taste, and decisions the project's authority reserves for them.

For studio applications, the running product is usually the best iteration surface. A separate mockup earns its cost only when it resolves a real design uncertainty that the running app cannot expose as cheaply. This removes a redundant artifact, not the visual review or proof contract.

Repair the narrowest authoritative layer

Find the generator or source that every affected agent actually consumes. Fix a code defect in code; fix a missing trigger in the skill; fix a malformed contract in its schema. Then verify through the consumer. Editing a rendered or installed copy creates a temporary fork, not a repair.

The payoff compounds: a correction to the shared application control plane changes the next agent's starting point. That is how an ongoing agent-operated system improves without making one human its permanent router.

04 同仓库更多 Skills

mvschwarz
mvschwarz/openrig-world

notes

Use when taking notes.

★ 16
mvschwarz
mvschwarz/openrig-world

software-for-agents

Use when older OpenRig instructions name software-for-agents or describe agent-in-the-middle or studio software under the former skill name.

★ 16
mvschwarz
mvschwarz/openrig-world

session-compaction-and-restore

Use when reasoning about what survives compaction/context-loss/restart/seat-refresh, designing a high-fidelity restore packet, or distinguishing native runtime resume vs fork vs artifact-backed mental-model rebuild. Covers the 4 failure modes that prevent honest restore (compacted seat drops hot potato; restore packet preserves details but loses product intent; runtime resume mistaken for handover or fork; rebuilt seat starts with stale instructions).

★ 16
mvschwarz
mvschwarz/openrig-world

seat-continuity-and-handover

Use when replacing a seat's occupant (rebuild/handover/swap), reasoning about stable-seat-identity vs fluid-occupant-identity, choosing an old-occupant disposition (retire/advise/shadow), or recording provenance for an occupant change. Two independent outcomes (continuityOutcome + seatBindingOutcome) and the 5 failure modes that prevent silent dishonesty.

★ 16
mvschwarz
mvschwarz/openrig-world

retiring-and-inheriting-a-seat

Use when you are a sitting agent near your context threshold (~85%) and a PLANNED seat transition is due — retire deliberately and hand your seat to a fresh successor primed from a packet, rather than let compaction degrade you. Covers the handover packet, the append-only lineage ledger (one row per tenure), physical-seat continuity, the do-not-over-inherit framing, the optional warm-handoff window, and wake-v0 (consulting a retired predecessor). NOT for unplanned compaction/crash recovery (session-compaction-and-restore / claude-compaction-restore) and NOT the seat-binding primitive mechanics (seat-continuity-and-handover).

★ 16
mvschwarz
mvschwarz/openrig-world

refocusing

Use when a long-running agent may have lost the product outcome, when context was compacted, when work has crossed a major boundary, or when a fresh path-based trace is needed before the next consequential action. Not for fresh-session orientation, waking an idle seat, or checkpoints.

★ 16