Your AI workforce is running on folklore. Put it on the approved playbook.
Each agent you deploy is taught by whoever set it up — no review, no record. The Agent Skills Library gives them all one governed playbook.
The skill sharing service for your AI workforce — authored by people, approved before it counts, followed by every agent.
| Rev | Origin | On the record |
|---|---|---|
| v1 Created | Written directly | author · time · change note |
| v2 Edited | Approved from a proposal | attributed to the submitter · reviewer on record |
| v3 Updated | Direct edit by a designated editor | versioned · attributed |
| v4 Restored → v2 | Recorded as a new version | history intact |
| v5 Proposed | Awaiting review | in the queue |
What are agent skills — and what is an agent skills library?
Agent skills are versioned, approved instruction packages an AI agent loads on demand: the task, the steps to do it your way, and the record of who approved them. An agent skills library is the one governed collection those packages live in, so every agent works from the same approved version.
A skill
One task, written down the way your company does it — what it is for, when it applies, and the steps. An agent loads it when that task comes up.
A skill file
The shape the market has settled on: one plain file per skill — often named SKILL.md — carrying a title, a short summary of when it applies, the steps, and tags.
An agent skills library
The approved collection those files live in: one source for people and every AI agent, with a review gate in front of it and full version history behind it.
Skills, tools and personas do different jobs
An AI agent skills library holds skills. Tools and personas are governed elsewhere, by different people, on a different cadence — which is why the three are worth keeping apart.
Scroll the comparison sideways
| What it is | What it changes | Who approves it | |
|---|---|---|---|
| Skill | A written playbook for one task | How the work gets done here | A designated reviewer, on the record |
| Tool | A capability an agent can call | What the agent can reach | Whoever grants the access |
| Persona | Standing framing for role and voice | How the answer reads | The team that owns the agent |
How the three differ in governance terms is set out on the agents, personas and skills page, and the wider picture of how agent work is coordinated is on the agentic AI page.
How a skill is evaluated before it is approved
A proposal reaches a designated reviewer, who weighs it against 5 plain criteria before it becomes what people and agents see by default.
After approval, agent use is recorded separately from human reads — the usage record of what your agent fleet relies on. That record is captured and retained today; reporting views over it are still being built.
You are hiring AI agents faster than you are governing them.
You spent years making software delivery trustworthy — versioned, reviewed, attributable. The instructions your AI agents actually follow have none of that discipline.
Scroll the comparison sideways
| Property | Your software delivery | Your agent instructions today |
|---|---|---|
| Version control | Governed Every change tracked with an author and a note; history is permanent. | Overwritten An edit replaces the last instruction — the old version is gone. |
| Review gate | Reviewed Changes pass review before they reach production. | No gate Anyone can rewrite what agents do, with no approval and no reviewer. |
| Shared source | Shared One repository the whole organization builds from. | Private Each team hand-crafts its own instructions; nothing is shared or reused. |
| Attribution | Attributed Every line traces to who wrote it and why. | Anonymous No author, no time — human and AI text indistinguishable. |
| Auditability | Auditable You can reconstruct who did what, when, and on what basis. | No record No record of which instruction an agent followed, or who changed it. |
Every new agent and every new team compounds the drift; there is no shared catalog to pull from.
Every departure before the library exists is know-how you never get back.
One approved playbook — for people and every agent.
A skill is a plain, structured playbook: how a task is done here. Authored once, approved once — then every agent works from the same version.
Consistency
The same task is performed the company’s way in one department exactly as it is in another.
Control
Staff propose; designated reviewers approve; approved content is what people and agents see by default.
Accountability
Every version carries who wrote it, when, and how it came to be.
Reversibility
A wrong change is undone by restoring an earlier version — and the restore is itself recorded.
Speed to capture
AI drafts a playbook from three plain questions; a person reviews it before it counts.
Sovereignty
It runs inside your own environment, behind your own sign-in — an asset you hold.
Published means approved — by construction.
Every change moves through one gate: proposed, reviewed, published, used, improved. Nothing reaches your agents without a designated authority behind it, and nothing is ever lost.
Draft
A person writes it — or AI drafts it from three plain questions. Never saved on its own.
Review
A reviewer approves, or rejects with a written reason. Each decision is recorded exactly once.
Publish
Into the approved library people and agents see by default.
Use
People read; agents fetch — each use recorded separately.
Improve
Edits and proposals re-enter the gate. History stays intact.
Any two versions, line by line.
Any earlier version — restored as a new version, never by rewriting history.
Every saved version records its origin; every AI drafting request is additionally on its own record.
Most staff propose into a queue. Designated editors — and an agent working under an editor’s access — may save directly, every change still versioned and attributed. Editor designation is administered with the platform operator today, deliberately.
Once published, an approved skill is simply something an orchestrator loads at the step that needs it — the mechanics of that are covered under agent orchestration. Teams that want the first set of playbooks written and the gate operated alongside their own reviewers can bring in Iternal’s AI agent development services.
Two doctrines run end to end.
History is never rewritten, and authority fails closed. Everything else — attribution, approvals, erasure, sign-in — is an expression of one of the two.
History is never rewritten
- Every change is a new version with a name, a time, and a note.
- A restore appends; nothing deletes the past.
- Reviewer identity and written reasons stay on the record.
- Every AI drafting request is separately recorded, with the settings and model that produced it.
Authority fails closed
- No verified identity, no access.
- An agent never exceeds the authority of the person who issued its credentials.
- Machine credentials are never administrators.
- One organization cannot see another — invisible, not merely forbidden.
- Locked-down deployments refuse service rather than degrade.
You see what the agent fleet actually relies on, separately from what people read.
Inside your walls
The library and its history live in your environment, behind your own sign-in.
The one path out
AI drafting goes only to the AI service you choose — a fully offline option keeps sealed environments self-contained.
Destruction is deliberate
Everyday removal is recoverable; permanent removal is a stricter, administrator-only act — and mistakes never leak internal detail.
The window is strategic, not technical.
Companies are hiring AI agents faster than they are governing them. Those who write the playbook now get compounding leverage; the rest run on folklore.
Compounding
Every skill written once improves every agent that follows. Models commoditize; your approved playbooks are knowledge only you hold.
Attrition
Your best procedures live in heads and private chat histories. Capture them while the people holding them are here.
Trust
Agents multiply the consequence of every unreviewed instruction. A governed playbook is what makes the next wave safe to adopt.
The perimeter rules and the fail-closed authority model above line up item by item with the AI agent security checklist, read from the reviewer’s side of the gate.
The Agent Skills Library is the playbook pillar of Iternal’s agentic suite — alongside Ultramemory for governed memory and the rest of the product line — each adoptable on its own. Where those pillars sit inside a wider agentic AI architecture is set out separately.
What is running now — and exactly how far it goes.
11 capabilities in the product today, stated plainly with the two statuses we actually use: ships and partial.
Scroll the register sideways for status
| Ref | Capability | Status |
|---|---|---|
| 01 | One approved, organization-wide library — with a draft → published → archived lifecycle; search shows approved content by default. | Ships |
| 02 | Full version history — every edit kept as a sealed, numbered version with author, time, change note, and origin; nothing overwritten. | Ships |
| 03 | Any two versions compared line by line | Ships |
| 04 | Restore any earlier version, recorded as a new version — today via the agent connection; the web app’s one-click restore button is still being added. | Partial |
| 05 | Proposal-and-review workflow — staff propose with notes; reviewers approve or reject with a written reason; each decision recorded exactly once. | Ships |
| 06 | AI-drafted playbooks from three plain questions — never saved without a human decision; every draft’s settings and model on the record. | Ships |
| 07 | Organization-tuned drafting — your context, your writing voice, your model choice, versioned like everything else. | Ships |
| 08 | Curated filing system with browse and search — the everyday controls in the web app today, the rest ready underneath. | Partial |
| 09 | Layered permissions with per-playbook sharing — honored end to end; the granting controls are still being surfaced. | Partial |
| 10 | Usage record — agent use logged separately from human reads; captured today, in-product reporting views still being built. | Partial |
| 11 | A standard connection point agents use to search, fetch, draft, and restore — connecting is configuration, not a build. | Ships |
A go/no-go signal in weeks, not quarters.
Three drills, each ending in something your own team verifies.
Their key playbooks, approved
Two or three teams author their first playbooks — by hand or AI-drafted from three questions — and a reviewer approves them in.
The control loop, exercised
Propose an edit; watch it queue, get approved — or rejected with the reason kept on the record — then restore an earlier version.
One agent, measured
Connect one agent and let it pull approved playbooks into its work.
The governance questions, answered plainly.
10 questions an AI-workforce deployment gets judged on — and what the product actually supports.
Q·01
If someone changes what our AI is told to do, can we see who, what, and why?
Yes. Every edit is saved as a new version stamped with its author, the time, and a change note — and each version records how it came to be: written directly, approved from a staff proposal, or restored from history. Every AI drafting request is separately recorded. Content is never overwritten, so what changed, who changed it, and why is always on file.
Q·02
Can we stop unreviewed instructions from reaching our agents?
Yes. Most staff propose changes into a review queue; only designated reviewers approve, and approved content is what people and agents see by default. Designated editors — and an agent operating under an editor’s access — can update the library directly, with every change still versioned and attributed. Editor designation is administered with the platform operator today, not self-serve — a deliberate control.
Q·03
Will every agent perform the same task the same way?
Yes. All agents draw from the same approved library, so a task is done the company’s way regardless of team, department, or which agent runs it. Improve the playbook once and every agent improves with it.
Q·04
Can we tell which playbooks our agents actually use?
Yes — every agent use is recorded distinctly from a person reading, building a usage record of what your agent fleet relies on. Today that record is captured and retained; in-product reporting views over it are still being built.
Q·05
If a change turns out to be wrong, can we undo it?
Yes. Any earlier version can be restored, and the restore is itself recorded as a new version — history is never rewritten. Today restores run through the agent connection; the web app’s one-click restore button is still being added. Only an administrator can remove a playbook outright.
Q·06
Can AI help us write these down without losing control?
Yes. AI drafts a playbook from three plain questions — the goal, the tools involved, the outcome — and the draft never enters the library on its own: a person reviews and saves it first in the normal flow. Every drafting request is recorded with the settings and model that produced it, so AI-assisted authoring never removes the human decision to publish.
Q·07
How is a skill evaluated before it is approved?
A proposal reaches a designated reviewer, who weighs it against five plain criteria: scope (one task per playbook), steps (each one followable by a person or an agent without asking a colleague), a stated finished result so the outcome can be judged rather than the intent, a named author and reviewer, and no duplication of a playbook already published. Approval publishes it as an attributed version; a rejection carries a written reason, and each decision is recorded exactly once.
Q·08
How do we keep an approved skill from going stale?
Improving a playbook is an edit and a review, not a rewrite: the proposal, the decision and every earlier version stay on the record, so a team can raise the standard without losing what came before. Agent use is recorded separately from human reads, building the usage record of which playbooks the fleet leans on — captured and retained today, with reporting views over it still being built. Retiring a playbook archives it, recoverably.
Q·09
How is a skill different from a tool or a persona?
A skill is the written playbook for one task — how the work is done here — approved by a designated reviewer and loaded by an agent when that task comes up. A tool is a capability the agent can call, governed by whoever grants the access. A persona is standing framing for role and voice, owned by the team running the agent. The Agent Skills Library governs the first of the three.
Q·10
Does our operating knowledge leave the building?
The library and its history live entirely inside your own environment, behind your own sign-in. The one configurable exception: AI-assisted drafting sends the drafting questions to whichever AI service you configure — including a fully offline option — and you control that choice.
Author it once. Approve it once. Every agent, the same playbook.
Turn how work gets done here into an owned, permanent asset — governed, versioned, and approved before the next wave of agents arrives. The library you build stays yours regardless of which AI providers you choose next.
Claims trace to the Agent Skills Library engineering record as of July 2026; partial statuses are stated exactly; figures labeled illustrative are concepts, not measurements.