agent-severance
Offboard an AI agent the way you'd offboard an employee — inventory what it knew and touched, export then purge its memory, revoke every credential …
它会碰到什么
这一栏是扫描器报的事实,不是结论。命中多不等于有毒(安全工具、规则库、示例脚本本来就会包含危险写法),命中少也不等于干净。它和你手上的凭据、文件、网络有什么关系,需要你自己看。
技能内容
Agent Severance Skill
Orgs learned employee offboarding the hard way: the contractor whose VPN worked
for a year after the contract, the shared password nobody rotated. Long-lived
agents recreate every one of those failure modes with worse logging — an agent
accumulates credentials, memory, integrations, scheduled jobs, and undocumented
responsibilities, and then one day it's "turned off" by deleting a chat window
while its API keys live on. This skill runs the severance properly: know what
it had, keep what's valuable, kill what's live, and hand over what it did.
What This Skill Produces
- An inventory: everything the agent could touch (credentials, tools, data
stores, channels), everything it knew (memory, context files, fine-tuning or
instructions), and everything it did on a schedule
- An access-revocation table with owner and verification step per row —
revoked isn't revoked until someone confirmed the key is dead
- A memory disposition record: exported / retained (where, why, how long) /
purged (how verified) — the part compliance will ask about in 2027
- A successor handover: the agent's actual duties, including the
undocumented ones users discovered, for whoever inherits them
Required Inputs
Ask for (if not already provided):
- The agent: platform, what it was for, how long it ran, who owned it
- Known integrations and credentials (then treat the list as incomplete on
principle — the inventory step hunts for the rest)
- Why it's being offboarded (vendor switch, pilot ended, incident, cost) —
incident-driven severance changes the order: revoke first, inventory second
- What must survive: memory worth exporting, workflows someone still needs
Process
- Inventory before touching anything (unless incident — then revoke
first). Hunt beyond the known list: API keys and OAuth grants · service
accounts · webhook URLs pointing at it · scheduled/cron jobs it ran ·
channels it posted in · data stores it read or wrote · other agents that
called it (the A2A dependencies nobody documented) · what its memory
contains, including personal data.
- Decide memory disposition per store, not wholesale. Export what has
value (decisions log, learned context) to an owned location; name a
retention owner and period for anything kept; purge the rest and record
how purged (vendor deletion request ≠ deleted — note what the vendor
actually promises). Personal data follows your privacy policy's deletion
rules, flagged explicitly.
- Revoke with verification. Every row gets: the credential, who revokes
it, and the test that proves it's dead (the call that now fails). Rotate
any shared secrets the agent ever held — its copy dying doesn't kill the
copies.
- Write the honest handover. What it was supposed to do, what it actually
did (ask its users — there are always undocumented duties), open threads
mid-flight, and the workflows that will silently break next Tuesday when it
stops.
- Announce the death. One message to the channels it served: it's gone,
here's who/what replaces it, here's where its exported memory lives.
Output Format
## Severance summary
[Agent, tenure, reason, severance owner, target date]
## Inventory
Access: [table — system, grant, discovered-how] · Knowledge: [stores + contents]
Duties: [scheduled + reactive + undocumented] · Dependents: [who/what calls it]
## Memory disposition
| Store | Export → where | Retain (owner, period) | Purge (method, verified how) |
## Revocation
| Credential/grant | Revoked by | Dead-key test | Status |
## Handover to successor
[Duties with enough detail to actually run them · open threads · will-break list]
## Announcement
[The message to its channels]
Quality Checks
- [ ] The inventory includes at least one category the user didn't mention —
scheduled jobs and agent-to-agent callers are the usual blind spots
- [ ] Every revocation row has a verification test, not just an action
- [ ] Shared secrets the agent held are rotated, not just revoked
- [ ] Memory disposition distinguishes vendor-promised deletion from verified
deletion, and flags personal data
- [ ] The handover names what breaks when the agent stops — if the answer is
"nothing", the duties inventory probably isn't done
Anti-Patterns
- [ ] Do not equate "deleted the chat/app" with offboarded — the checklist
exists because credentials outlive interfaces
- [ ] Do not purge memory before the export decision — severance is
irreversible in exactly one direction
- [ ] Do not skip the users interview; the undocumented duties are the ones
that page someone at 2am after shutdown
- [ ] Do not write this as a vendor grudge document — it's an operational
record compliance and the successor will both read
Related
[[agent-hiring-panel]] is the front door this is the back door of;
[[context-bankruptcy]] when the agent stays but its memory shouldn't;
[[agent-incident-postmortem]] if an incident triggered this.
想直接用这个技能?
本站把开放许可(MIT / Apache 等)的技能按仓库打包整理到网盘,点一下转存到你自己的网盘,不用一个个从 GitHub 拉。许可未声明的技能只给原始仓库链接,不打包。
同名技能的其他版本
有 3 个不同仓库或目录里都有叫 agent-severance 的技能。它们内容并不相同,别混用:
- mohitagw15856/pm-claude-skills — Offboard an AI agent the way you'd offboard an employee — inventory what it knew and touch
- mohitagw15856/pm-claude-skills — Summarise what Agent Severance does in one line. Use when asked to [trigger phrases the us