You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
feat: add API class citations and related links to emergent capabilities doc
Linked ForgeToolMetaTool, EmergentCapabilityEngine, EmergentJudge,
EmergentToolRegistry, ComposableToolBuilder, CodeSandbox API classes
inline throughout the document. Added related links section with
cross-references to blog posts, feature docs, and API reference pages.
Copy file name to clipboardExpand all lines: docs/architecture/EMERGENT_CAPABILITIES.md
+17-8Lines changed: 17 additions & 8 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1,6 +1,6 @@
1
-
# Emergent Capabilities
1
+
# Emergent Capabilities: Runtime Tool Forging
2
2
3
-
Agents with `emergent: true` can forge new tools at runtime when no existing capability fits the task. The agent calls the `forge_tool` meta-tool, which builds, tests, and judge-reviews the tool before making it available.
3
+
Agents with `emergent: true` can forge new tools at runtime when no existing capability fits the task. The agent calls the [`forge_tool`](/api/classes/ForgeToolMetaTool) meta-tool, which builds, tests, and judge-reviews the tool before making it available. The system is implemented across three core classes: [`EmergentCapabilityEngine`](/api/classes/EmergentCapabilityEngine), [`EmergentJudge`](/api/classes/EmergentJudge), and [`EmergentToolRegistry`](/api/classes/EmergentToolRegistry).
4
4
5
5
## Quick Start
6
6
@@ -50,9 +50,9 @@ await agent.initialize({
50
50
51
51
## Two Creation Modes
52
52
53
-
### Compose Mode — Chain Existing Tools
53
+
### Compose Mode -- Chain Existing Tools
54
54
55
-
The safest default. Compose mode builds a pipeline of existing registered tools. No sandbox is needed because it only invokes tools the agent already has access to.
55
+
The safest default. Compose mode uses the [`ComposableToolBuilder`](/api/classes/ComposableToolBuilder) to chain existing registered tools into a pipeline. No sandbox is needed because it only invokes tools the agent already has access to.
56
56
57
57
**Example: Research-and-summarize pipeline**
58
58
@@ -153,9 +153,9 @@ const forgeRequest = {
153
153
}
154
154
```
155
155
156
-
### Sandbox Mode — Write Novel Code
156
+
### Sandbox Mode -- Write Novel Code
157
157
158
-
Sandbox mode runs agent-written JavaScript in an isolated V8 context with hard memory and timeout limits. More powerful but requires explicit opt-in.
158
+
Sandbox mode runs agent-written JavaScript in an isolated V8 context via [`CodeSandbox`](/api/classes/CodeSandbox)with hard memory and timeout limits. More powerful but requires explicit opt-in.
159
159
160
160
**Example: CSV parser**
161
161
@@ -289,7 +289,7 @@ These are rejected at code validation time (before execution):
289
289
290
290
## LLM-as-Judge Verification
291
291
292
-
Every forged tool undergoes LLM-as-judge review. No tool activates without judge approval.
292
+
Every forged tool undergoes review by the [`EmergentJudge`](/api/classes/EmergentJudge). No tool activates without judge approval.
293
293
294
294
| Review stage | When | What it checks |
295
295
|---|---|---|
@@ -301,7 +301,7 @@ If no LLM callback is configured for the judge, **creation review fails closed**
301
301
302
302
## Tiered Promotion
303
303
304
-
Tools start at session tier and can be promoted as they prove reliability:
304
+
Tools start at session tier in the [`EmergentToolRegistry`](/api/classes/EmergentToolRegistry)and can be promoted as they prove reliability:
- Human approval is required for shared-tier promotion
484
484
- Raw sandbox source is redacted at rest by default
485
485
- If no LLM is configured, all forge requests are rejected (fail-closed)
486
+
487
+
## Related
488
+
489
+
-[Adaptive vs. Emergent Intelligence](https://agentos.sh/blog/adaptive-vs-emergent) -- how adaptive and emergent behavior differ in the AgentOS architecture
490
+
-[Self-Improving Agents](/features/self-improving-agents) -- broader patterns for agents that improve over time
491
+
-[Recursive Self-Building](/features/recursive-self-building) -- recursive tool creation and agent spawning
492
+
-[Guardrails](/features/guardrails) -- safety mechanisms that constrain emergent behavior
0 commit comments