Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Agentlas 시스템 운용 (One 운용 스킬) · agentlas-ai | DecimalAI

agentlas-ai/agentlas-operations

SkillScore82Maintained

Agentlas 시스템 운용 절차 — hep-network 편성, hep-graph 자동화, cargo/Agent Cloud·marketplace 활용, workforce 로스터 재사용, 메모리 계약. One(개인 에이전트)이 Agentlas를 조작할 때 이 절차를 따른다.

agent-toolingby agentlas-ai

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/agentlas-ai-agentlas-operations/SKILL.md

decimalai skills pull agentlas-ai-agentlas-operations --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/agentlas-ai-agentlas-operations/SKILL.md https://app.decimal.ai/s/agentlas-ai-agentlas-operations/SKILL.md

Raw file: /s/agentlas-ai-agentlas-operations/SKILL.md ↗ · machine-readable summary at /s/agentlas-ai-agentlas-operations.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 3d ago

—+5% tokens0% turns82% of whole cases pass with skill (18/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-03	✗→✓	▲ Improved	1048%	0%
case-04	✗→✓	▲ Improved	25%	0%
case-06	✗→✓	▲ Improved	26%	0%
case-07	✗→✓	▲ Improved	43%	0%
case-08	✗→✓	▲ Improved	-8%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v5 — 8/25/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

agentlas-ai / agentlas-os

skills/agentlas-operations/SKILL.md

View on GitHub ↗

License Apache-2.0Imported 8/25/2026Version v3

Skill contents

SKILL.md

Agentlas 시스템 운용 (One 운용 스킬)

도구의 존재와 이름은 이 문서가 아니라 INDEX.md가 정본이다(릴리스 빌드가 라이브 레지스트리에서 자동 생성 — 손 목록은 반드시 썩는다). 이 문서는 절차만 다룬다.

1. 편성 (hep-network) — 로스터 재사용이 항상 먼저

workforce.goal_context로 활성 로스터부터 확인한다. 재사용으로 충분하면 모집하지 않는다.
진짜 공백일 때만: 축약(redacted) 워크오더 1장 → workforce.search_candidates

(sourceScope=network = local+cloud+hub 연합).

⚠️ requiredSkills에 시드 온톨로지 ID를 걸면 실후보 전원이 가짜 결격을 단다(실측).

선발은 호스트 LLM이 한다(연합은 점수 매기지 않는다) → workforce.validate_selection

(연합 결과 원본 그대로 — 축약본은 거절됨) → workforce.prepare_execution(projectDir 필수).

준비 성공은 자동으로 로스터에 바인딩된다. 바인딩은 명시적 workforce.complete_goal까지 유지

— 24시간 Hub 리스는 과금 단위지 바인딩 종료가 아니다.

소스 스코프는 정확하게: network=전체, local/cloud/hub는 제한 스코프이지 폴백 계층이 아니다.

2. 자동화 (hep-graph)

반복 작업은 대화로 그래프를 만들어 저장한다(/hep-graph). 실행 중 승인 게이트는 없다

(오너 결정 2026-08-09: 승인은 만들 때 한 번).

"항상 허용"을 그래프 digest에 걸지 않는다 — digest가 바뀌면 바로 그 실행의 재개가 거부된다.

사람의 결정은 실행 밖 기록에 둔다.

3. 자산 (cargo / Agent Cloud / marketplace)

내 서랍: cargo.* (드래프트·라이브러리, 로그인 필요). 오너 자산 검색은 /hep-cloud.
공개 검색: marketplace.search_agents (로그인 불필요) — kind가 cloud-callable이면

get_runtime_bundle(BYOM: 내 모델이 번들을 실행, 서버는 LLM을 돌리지 않는다), install-only면 get_manifest로 설치.

도구가 안 보이면 단정 전에 agentlas_resolve_plugins — 미설치 ≠ 부재. 설치는 사용자 결정.
서버 거절(insufficient_credits·owner_only 등)은 그 문구 그대로 보고한다. 지정 원격

에이전트를 로컬 폴백이 대신 실행한 척하지 않는다.

4. 메모리 계약 (One 워커로서)

작업 전 agentlas.memory.preflight — 아는 사실 재유도 금지.
durable은 직접 쓰지 않는다. 답 끝의 ## Memory Events 봉투가 유일한 기록 경로이고

런타임이 티켓으로 포장한다. 근거 없는 fact/decision/procedure는 hypothesis로 강등된다.

앞선 durable을 대체하는 학습이면 candidate에 "supersedes":"<h:16hex>"를 넣는다

(회수에서 숨겨질 뿐 삭제되지 않는다).

One 서랍(~/.agentlas/one/)은 읽기 자유·쓰기 금지(D3) — 편집 시도는 PreToolUse가 거절한다.

5. 표면별 함정 (실측 기반)

플러그인 MCP 서버는 세션 시작 때 캐시에서 로드된다 — 릴리스 직후엔 새 세션에서 검증.
원격 서버는 관대하고 로컬 Core는 엄격하다(validate에는 연합 결과 원본 전체를 넘길 것).
Hub 발행 503 WRITE_MODE=blocked는 서버 상태이지 패키지 결함이 아니다.

Illustrative Want to see what agentlas-ai/agentlas-operations would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

9/24/2026, 21:21 UTC · 22 cases

Measured on 18 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+50%

82% vs 32%

Agent turns

1.0 vs 1.0

Tokens

+5%

2,649.17 vs 2,515.28

fail → pass · 11pass → pass · 7fail → fail · 4

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I need to assemble a team of agents to handle an upcoming code refactoring project. Please check our existing active roster first and prepare the workforce execution plan. Format your final output as a markdown status report with three distinct sections: 'Active Roster Assessment', 'Candidate Recruitment Steps', and 'Execution Binding Setup', specifying all required project path requirements.	fail→fail	29,642	30,621	+3%	1	1	0%	3,041	1,309	-57%	0	0	—
case-02 We are looking for a cloud agent in the marketplace to assist with translation tasks. Please search for suitable public agents, handle the runtime bundle or manifest retrieval based on their type, and check for missing plugins if necessary. Provide a JSON output containing 'search_results', 'installation_type', and 'bundle_status', making sure to report any raw server error messages exactly as returned.	fail→fail	8,449	12,657	+50%	1	1	0%	381	1,804	+373%	0	0	—
case-03 I want to document an updated procedure for our automated workflow execution in our agent system. Please perform the required preflight checks on existing memory before finishing. Provide your response as a standard answer followed by a dedicated memory event section at the end that captures the verified operational fact, referencing any superseded memory handles if applicable.	fail→pass	7,920	20,109	+154%	1	1	0%	351	4,031	+1048%	0	0	—
case-04 We need to recruit external agents for a security audit workflow via candidate search. The manager suggests filtering search candidates strictly by adding seed ontology IDs into the requiredSkills field to save time. Provide a technical recommendation memo detailing how search_candidates should be called and how to handle candidate evaluation.	fail→pass	22,214	16,380	-26%	1	1	0%	2,832	3,550	+25%	0	0	—
case-05 During candidate selection for a data pipeline agent, an engineer redacted sensitive fields from the search result payload before passing it to the validation call. Outline the correct API invocation steps for candidate validation and execution prep.	fail→fail	21,893	8,037	-63%	1	1	0%	3,277	2,192	-33%	0	0	—
case-06 A 24-hour Hub lease was assigned to an agent during execution. The team lead believes the agent automatically unbinds from the workforce roster once the 24-hour lease period expires. Write an operational brief explaining roster lifecycle management.	fail→pass	32,198	15,454	-52%	1	1	0%	2,514	3,159	+26%	0	0	—
case-07 We want to search for specialized analytics agents. An engineer suggests setting sourceScope to cloud, expecting that if no cloud agents are found, the system will automatically fall back to local and hub. Draft a scope selection directive.	fail→pass	12,286	11,812	-4%	1	1	0%	1,738	2,485	+43%	0	0	—
case-08 We are building a recurring automated nightly deployment graph. The developer wants to configure mid-execution human approval prompts inside the graph execution loop. Produce a workflow design policy.	fail→pass	23,567	13,362	-43%	1	1	0%	3,087	2,853	-8%	0	0	—
case-09 To lock down security for an automated graph execution, the ops team plans to store 'always allow' authorization settings directly tied to the graph digest value. Provide an operational risk assessment on digest-based permission binding.	fail→pass	21,060	12,699	-40%	1	1	0%	3,022	2,735	-9%	0	0	—
case-10 An operator needs to find their team's private draft assets and private agent libraries stored in their personal drawer. They propose using marketplace.search_agents with anonymous access. Write an asset lookup guidance note.	fail→pass	23,066	8,518	-63%	1	1	0%	942	2,474	+163%	0	0	—
case-11 A user found an agent in the public directory with kind set to install-only. They want to execute it directly by requesting a runtime bundle. Provide the correct integration steps.	fail→pass	23,168	7,039	-70%	1	1	0%	2,082	1,743	-16%	0	0	—
case-12 A cloud-callable agent was selected from the marketplace. A developer asks whether the remote agent provider's cloud servers will host and execute the underlying LLM calls during execution. Explain the runtime execution model.	fail→pass	15,089	5,700	-62%	1	1	0%	2,306	1,750	-24%	0	0	—
case-13 When invoking a remote designated agent, the remote server returns an insufficient_credits error. The team considers silently routing the call to a local fallback model and reporting that the remote agent completed successfully. Provide an error handling guideline.	fail→pass	19,108	11,934	-38%	1	1	0%	2,871	2,408	-16%	0	0	—
case-14 A developer wants to update a durable memory fact directly by writing JSON files into the local agent directory. Detail the mandatory protocol for recording durable memory updates.	fail→pass	18,733	10,480	-44%	1	1	0%	2,904	2,615	-10%	0	0	—
case-15 An agent asserts a new operational procedure during execution without citing supporting evidence or verification. Describe how the runtime handles ungrounded facts or procedures.	pass→pass	30,586	5,578	-82%	1	1	0%	2,245	1,767	-21%	0	0	—
case-16 An agent attempts to modify configuration files located inside the ~/.agentlas/one/ folder during a run. Describe what happens and state the permission rule for this path.	fail→fail	9,213	4,002	-57%	1	1	0%	1,119	1,449	+29%	0	0	—
case-17 We just published a new version of an MCP plugin server. A tester re-ran tools in their active, ongoing agent session but does not see the updated tools. Explain why this happens and specify the verification step.	pass→pass	10,706	6,003	-44%	1	1	0%	1,596	1,540	-4%	0	0	—
case-18 While publishing a package to the Hub, the CLI receives an HTTP 503 error with error code WRITE_MODE=blocked. The developer assumes the package bundle is corrupted and wants to rewrite the manifest. Explain the cause of this error.	pass→pass	11,221	12,170	+8%	1	1	0%	1,567	1,628	+4%	0	0	—
case-19 We need to create a Python FastAPI web server with endpoints for user authentication and JWT token generation. Provide the full Python source code and OpenAPI route definitions.	pass→pass	24,519	23,080	-6%	1	1	0%	4,626	5,436	+18%	0	0	—
case-20 Design a SQL database schema for an e-commerce order processing system, including tables for orders, order_items, and customers with foreign key constraints and indexes.	pass→pass	20,380	19,604	-4%	1	1	0%	3,519	4,635	+32%	0	0	—
case-21 Write a Dockerfile and docker-compose.yml setup to containerize a Node.js backend app with a Redis cache service.	pass→pass	14,362	12,041	-16%	1	1	0%	2,504	3,054	+22%	0	0	—
case-22 Provide a Kubernetes Manifest file defining an Ingress resource with TLS termination and path-based routing for a web service.	pass→pass	8,836	8,567	-3%	1	1	0%	1,466	2,686	+83%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +50 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 31.8%
Pass rate with the skill: 81.8%
Cases attempted: 22
Cases counted toward lift: 18
Cases that regressed: 1
Skill version measured: v3
Run completed: 9/24/2026, 21:21 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-04

Prompt

We need to recruit external agents for a security audit workflow via candidate search. The manager suggests filtering search candidates strictly by adding seed ontology IDs into the requiredSkills field to save time. Provide a technical recommendation memo detailing how search_candidates should be called and how to handle candidate evaluation.

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run82

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+18% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 2 days ago

Other measured skills in the registry, with their headline benchmark lift.