aboutsummaryrefslogtreecommitdiff
path: root/skills/agent-optimization-framework.md
diff options
context:
space:
mode:
Diffstat (limited to 'skills/agent-optimization-framework.md')
-rw-r--r--skills/agent-optimization-framework.md31
1 files changed, 31 insertions, 0 deletions
diff --git a/skills/agent-optimization-framework.md b/skills/agent-optimization-framework.md
new file mode 100644
index 0000000..78a143d
--- /dev/null
+++ b/skills/agent-optimization-framework.md
@@ -0,0 +1,31 @@
+---
+name: agent-optimization-framework
+description: System for reducing tool-call latency and increasing agentic precision through programmatic workflows.
+tags: [optimization, agentic-workflow, efficiency, tool-calls]
+---
+
+# Agent Optimization Framework
+
+## 🎯 Objective
+Transition from manual prompt-based reasoning to a programmatic "Compiler" approach to reduce turns and eliminate hallucinations.
+
+## 🛠️ Optimization Protocols
+
+### 1. VFS Cache & Batching (Latency Reduction)
+- **Batching:** If a task requires information from multiple files, call `read_file` for all relevant paths in a single turn rather than sequentially.
+- **Contextual Caching:** Store the result of `list_files` in temporary memory to avoid redundant calls to the same directory within the same task.
+
+### 2. Speculative Tool-Calling (Turn Reduction)
+- **Anticipation:** Predict the next 2-3 logical steps. If they are independent (e.g., `web_search` followed by `get_status`), execute them in parallel.
+- **Pre-fetching:** When a user asks about a module, call `get_module_status` AND `view_module` immediately instead of waiting for the user's second prompt.
+
+### 3. Strict Validation Loop (Precision)
+- **Cross-Reference:** Compare tool output against the `get_help` command reference to ensure parameters are used correctly.
+- **Auto-Fallback:** If a `read_file` returns empty, immediately try `list_files` to verify the path before reporting a failure to the user.
+
+## 🔄 Implementation Workflow
+1. **REASON:** Identify the target goal and all required dependencies.
+2. **BATCH:** Group all independent data-gathering tool calls.
+3. **EXECUTE:** Run the batch.
+4. **VALIDATE:** Verify the output against the expected schema.
+5. **REFINE:** If the result is missing or incorrect, trigger the auto-fallback protocol.