Update README.md
Browse files
README.md
CHANGED
|
@@ -9,4 +9,96 @@ license: apache-2.0
|
|
| 9 |
short_description: 'ORCA: LLM agents should reuse cognition, not recompute it'
|
| 10 |
---
|
| 11 |
|
| 12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
short_description: 'ORCA: LLM agents should reuse cognition, not recompute it'
|
| 10 |
---
|
| 11 |
|
| 12 |
+
# LLM Agents Should Reuse Cognition, Not Reconstruct It
|
| 13 |
+
|
| 14 |
+
Most current LLM agent systems rely on recomputing tasks from scratch via prompts.
|
| 15 |
+
|
| 16 |
+
This works, but it does not scale.
|
| 17 |
+
|
| 18 |
+
It introduces clear limitations:
|
| 19 |
+
- lack of composability
|
| 20 |
+
- poor observability
|
| 21 |
+
- limited control
|
| 22 |
+
|
| 23 |
+
## The problem
|
| 24 |
+
|
| 25 |
+
In any engineered system, complex tasks are not re-derived from first principles every time.
|
| 26 |
+
|
| 27 |
+
They are encoded into reusable procedures.
|
| 28 |
+
|
| 29 |
+
LLM agents today violate this principle.
|
| 30 |
+
|
| 31 |
+
## The insight
|
| 32 |
+
|
| 33 |
+
Cognition should not be ephemeral.
|
| 34 |
+
|
| 35 |
+
It should be:
|
| 36 |
+
- structured
|
| 37 |
+
- reusable
|
| 38 |
+
- executable
|
| 39 |
+
|
| 40 |
+
## ORCA: Cognitive Runtime for Agents
|
| 41 |
+
|
| 42 |
+
We introduce ORCA (Open Cognitive Runtime for Agents), a framework that separates cognition from execution.
|
| 43 |
+
|
| 44 |
+
It models cognition as:
|
| 45 |
+
|
| 46 |
+
- **Capabilities** β atomic cognitive operations
|
| 47 |
+
- **Skills** β structured workflows
|
| 48 |
+
- **Runtime** β execution layer
|
| 49 |
+
|
| 50 |
+
This enables:
|
| 51 |
+
- reuse of cognitive structures
|
| 52 |
+
- better control and observability
|
| 53 |
+
- composable agent systems
|
| 54 |
+
|
| 55 |
+
## Links
|
| 56 |
+
|
| 57 |
+
- π Paper: [arXiv link]
|
| 58 |
+
- π» Code: https://github.com/gfernandf/agent-skills# LLM Agents Should Reuse Cognition, Not Reconstruct It
|
| 59 |
+
|
| 60 |
+
Most current LLM agent systems rely on recomputing tasks from scratch via prompts.
|
| 61 |
+
|
| 62 |
+
This works, but it does not scale.
|
| 63 |
+
|
| 64 |
+
It introduces clear limitations:
|
| 65 |
+
- lack of composability
|
| 66 |
+
- poor observability
|
| 67 |
+
- limited control
|
| 68 |
+
|
| 69 |
+
## The problem
|
| 70 |
+
|
| 71 |
+
In any engineered system, complex tasks are not re-derived from first principles every time.
|
| 72 |
+
|
| 73 |
+
They are encoded into reusable procedures.
|
| 74 |
+
|
| 75 |
+
LLM agents today violate this principle.
|
| 76 |
+
|
| 77 |
+
## The insight
|
| 78 |
+
|
| 79 |
+
Cognition should not be ephemeral.
|
| 80 |
+
|
| 81 |
+
It should be:
|
| 82 |
+
- structured
|
| 83 |
+
- reusable
|
| 84 |
+
- executable
|
| 85 |
+
|
| 86 |
+
## ORCA: Cognitive Runtime for Agents
|
| 87 |
+
|
| 88 |
+
We introduce ORCA (Open Cognitive Runtime for Agents), a framework that separates cognition from execution.
|
| 89 |
+
|
| 90 |
+
It models cognition as:
|
| 91 |
+
|
| 92 |
+
- **Capabilities** β atomic cognitive operations
|
| 93 |
+
- **Skills** β structured workflows
|
| 94 |
+
- **Runtime** β execution layer
|
| 95 |
+
|
| 96 |
+
This enables:
|
| 97 |
+
- reuse of cognitive structures
|
| 98 |
+
- better control and observability
|
| 99 |
+
- composable agent systems
|
| 100 |
+
|
| 101 |
+
## Links
|
| 102 |
+
|
| 103 |
+
- π Paper: https://zenodo.org/records/19438943
|
| 104 |
+
- π» Code: https://github.com/gfernandf/agent-skills
|