နားလည်ထားရမယ့် အချက်
LLM တစ်ခုတည်းက function တစ်ခုပါ — text ဝင်၊ text ထွက်၊ ဒီကြောင့် ကမ္ဘာမှာ ဘာမှ မပြောင်းလဲပါဘူး။ Agent တစ်ခုက LLM တူတူကို ထပ်ဆောင်း piece လေးခုနဲ့ wrap လုပ်ပါတယ်
- Memory — step တွေတစ်လျှောက် ဆောင်ထားတဲ့ state
- Tools — file ဖတ်တာ ဒါမှမဟုတ် API ခေါ်တာလို model က invoke လုပ်နိုင်တဲ့ function
- Planning — goal တစ်ခုကို step တွေအဖြစ် ခွဲတာ
- Execution — ဒီ step တွေကို တကယ် run ပြီး ရလဒ်ကို ပြန်ပို့တာ
LLM ကိုယ်တိုင်ကတော့ မပြောင်းလဲပါဘူး — ပြောင်းလဲသွားတာက ၎င်းရဲ့ text output က real side effect တွေ ဖြစ်စေနိုင်တာပါ။
User request
User က request ပို့ပါတယ်။
LLM ဆုံးဖြတ်ချက်
LLM က available tool list နဲ့ ၎င်းတို့ရဲ့ description ကို ရရှိထားပြီး၊ ဖြေဆိုဖို့ tool တစ်ခု ခေါ်ဖို့ လိုအပ်လား ဆုံးဖြတ်ပါတယ်၊ လိုအပ်ရင် plain-text အဖြေအစား structured call (tool name plus arguments) တစ်ခု ထုတ်ပေးပါတယ်။
Program execute လုပ်ခြင်း
ဒီ tool ကို တကယ် execute လုပ်ပေးတာက ပတ်ဝန်းကျင် program က ဖြစ်ပါတယ်၊ model က မဟုတ်ပါဘူး — LLM မှာ သင့် filesystem, network, ဒါမှမဟုတ် shell ကို တိုက်ရိုက် access လုပ်ခွင့် လုံးဝမရှိလို့ပါ၊ ၎င်းက ဖြစ်စေချင်တာကို ဖော်ပြတဲ့ text ကိုပဲ ထုတ်ပေးနိုင်ပါတယ်။
Result ပြန်ပို့ခြင်း
Tool ရဲ့ result ကို conversation ထဲ ပြန်ထည့်ပေးပါတယ်၊ LLM က tool နောက်တစ်ခု ခေါ်မလား၊ ဒါမှမဟုတ် final answer ထုတ်ပေးမလား ဆုံးဖြတ်ပါတယ်။
Model က ၎င်းကိုယ်ပိုင် (တခါတရံ မှားနိုင်၊ တခါတရံ manipulate ခံရနိုင်တဲ့) reasoning ပေါ်မူတည်ပြီး tool ဘယ်ခေါ်မလဲ ဆုံးဖြတ်လို့ safety ဟာ model ရဲ့ judgment ထဲမှာ မဟုတ်ဘဲ ၎င်းပတ်ဝန်းကျင်ရှိ execution layer ထဲမှာ ရှိရမှာပါ။
- Least-privilege access — task တစ်ခုအတွက် လိုအပ်တဲ့ tool တိကျတိကျပဲ ခွင့်ပြု၊ ကျယ်ပြန့်တာ မလုပ်ဘူး
- Sandboxed workspace — agent မထွက်နိုင်တဲ့ directory ဒါမှမဟုတ် container
- Default အနေနဲ့ read-only tool၊ write/delete access ကို opt-in ဖြစ်စေခြင်း
- Open-ended shell အစား ခွင့်ပြုထားတဲ့ tool explicit allowlist
- Execute မလုပ်ခင် tool argument validation
- Call တိုင်းရဲ့ audit logging
- Timeout နှင့် network restriction
File ဒါမှမဟုတ် system access ကန့်သတ်ချက်မရှိတဲ့ agent တစ်ခုဟာ ပိုစွမ်းရည်ရှိတဲ့ assistant တစ်ခု မဟုတ်ပါဘူး — standing risk တစ်ခုပါ။
- Agent
- LLM တစ်ခုကို memory, tools, planning, execution layer တို့နှင့် ပေါင်းစပ်ထားတာပါ၊ text response တစ်ခုတည်း ထုတ်ပေးရုံမက real side effect ရှိတဲ့ multi-step task တွေကို လုပ်ဆောင်နိုင်ပါတယ်။
- Tool Calling
- LLM တစ်ခုက function အတိအကျတစ်ခုကို argument အတိအကျတွေနဲ့ run ဖို့ structured request တစ်ခု ထုတ်ပေးတဲ့ mechanism ပါ၊ ပတ်ဝန်းကျင် program က execute လုပ်ပြီး result ကို conversation ထဲ ပြန်ထည့်ပေးပါတယ်။
THE TOOL-CALLING LOOP
---------------------
THE TOOL-CALLING LOOP
------------------------
user request
|
v
LLM -----------------------------+
| decides a tool is needed |
v |
tool call (name + arguments) |
| |
v |
[ CONFIRMATION GATE ] <-- required for destructive tools
| (blocked until user confirms)
v
tool executes (outside the model)
|
v
result returned to LLM -------------+
|
v
LLM responds to user (or calls another tool)လက်တွေ့ scenario နဲ့ ချိတ်ကြည့်မယ်
အောက်က code ဥပမာက real agent framework တိုင်းက တစ်နည်းနည်းနဲ့ implement လုပ်ထားတဲ့ pattern ရဲ့ minimal version တစ်ခုပါ — execute မလုပ်ခင်၊ execute လုပ်ပြီးမှ မဟုတ်ဘဲ request လုပ်ထားတဲ့ tool call ကို run ခွင့်ပြုမလား ဆုံးဖြတ်ပေးတဲ့ lookup table တစ်ခုပါ။
ပထမ၊ default-deny — allowlist ထဲ လုံးဝမပါတဲ့ tool name တစ်ခုက safe ဖြစ်မဖြစ် system က မှန်းဆကြည့်မယ့်အစား တိုက်ရိုက် reject ခံရပါတယ် — ဥပမာထဲက `send_email` က ဘယ်လို argument ပါလာပါစေ grant ခံရဖူးခြင်း မရှိလို့ fail ဖြစ်ပါတယ်။
ဒုတိယ၊ သီးခြား destructive flag တစ်ခု — state ကို ပြောင်း/ဖျက်တဲ့ tool တွေ (ဒီမှာ `delete_file` နဲ့ `run_shell_command`) က allowlist ထဲပါတာအပြင် explicit `confirmed=True` ထပ်လိုအပ်ပါတယ်၊ read access နဲ့ write/delete access ကို bar တစ်ခုတည်း ဘယ်တော့မှ မမျှဝေသင့်ဘူးဆိုတဲ့ idea ကို model လုပ်ပြထားတာပါ။
Real implementation တစ်ခုက agent လုပ်တဲ့ tool call တိုင်းမလုပ်ခင် ဒါကို check ရမှာပါ၊ ဆုံးဖြတ်ချက်ကို ဘယ်လိုပဲဖြစ်ဖြစ် log ချရမှာပါ၊ `confirmed=True` ကို human (ဒါမှမဟုတ် human ကိုယ်တိုင် explicit သတ်မှတ်ထားတဲ့ policy) ပဲ ပေးနိုင်တာလို့ သဘောထားရမှာပါ။
Model ကို ကိုယ့်ကိုကိုယ် Confirm လုပ်ခွင့် ဘယ်တော့မှ မပေးပါနှင့်
LLM က ၎င်းရဲ့ tool-call argument ထဲ ထည့်ပြီး `confirmed=True` ကို ကိုယ့်ကိုကိုယ် confirm လုပ်လို့ ဘယ်တော့မှ မရပါဘူး — ဒါလုပ်ခွင့်ပေးရင် manipulate ခံထားရတဲ့ model တစ်ခုက ၎င်းရဲ့ destructive action ကို ကိုယ်တိုင် confirm လုပ်နိုင်စေမှာပါ။
ကန့်သတ်ချက်မရှိတဲ့ Agent Access ဟာ Feature မဟုတ်၊ Risk တစ်ခုပါ
Allowlist မရှိ၊ sandboxing မရှိ၊ confirmation step မရှိဘဲ file တွေကို ဖတ်၊ ရေး၊ ဖျက်နိုင်၊ shell command မည်သည့်အရာမဆို run နိုင်တဲ့ agent တစ်ခုဟာ ပိုစွမ်းရည်ရှိတဲ့ assistant တစ်ခု မဟုတ်ပါဘူး — ကြည့်ရှုသူမရှိဘဲ၊ တခါတရံ မှားနေတဲ့ script တစ်ခုကို သင့် system အပေါ် အပြည့်အဝ access ပေးလိုက်တာနဲ့ ညီမျှပါတယ်။ Destructive capability တိုင်းကို opt-in ဖြစ်ပြီး ရှင်းရှင်းလင်းလင်း scope လုပ်ထားရမယ့်အရာအနေနဲ့ သဘောထားပါ၊ default အနေနဲ့ ဘယ်တော့မှ မထားပါနဲ့။
အတူတူ စမ်းရေးကြည့်မယ်
ALLOWED_TOOLS = {
"read_file": {"destructive": False},
"list_directory": {"destructive": False},
"web_search": {"destructive": False},
"delete_file": {"destructive": True},
"run_shell_command": {"destructive": True},
}
def check_permission(tool_name, allowed_tools, confirmed=False):
if tool_name not in allowed_tools:
return False, "not in allowlist"
if allowed_tools[tool_name]["destructive"] and not confirmed:
return False, "destructive tool requires confirmed=True"
return True, "permitted"
calls = [
("read_file", False),
("send_email", False), # not in allowlist at all
("delete_file", False), # destructive, no confirmation
("delete_file", True), # destructive, confirmed
]
for tool_name, confirmed in calls:
ok, reason = check_permission(tool_name, ALLOWED_TOOLS, confirmed=confirmed)
print(f"tool={tool_name!r} confirmed={confirmed} -> allowed={ok} ({reason})")Example call လေးခုကို check_permission ထဲ run လိုက်ရာ -
tool='read_file' confirmed=False -> allowed=True (permitted)
tool='send_email' confirmed=False -> allowed=False (not in allowlist)
tool='delete_file' confirmed=False -> allowed=False (destructive tool requires confirmed=True)
tool='delete_file' confirmed=True -> allowed=True (permitted)
read_file က allowlist ထဲပါပြီး destructive မဟုတ်လို့ ချက်ချင်း allow ခံရပါတယ်။ send_email က allowlist ထဲ လုံးဝမပါလို့ deny ခံရတာပါ — ၎င်းရဲ့ destructive flag ကို check တောင် မလုပ်ပါဘူး။ delete_file က confirmation မပါဘဲ deny ခံရပြီး confirmation ပါလာရင် allow ခံရပါတယ်၊ layer နှစ်ခု check (allowlist membership, ပြီးရင် destructive tool အတွက် သီးခြား confirmation requirement) design အတိုင်း အလုပ်လုပ်နေတာကို ပြသပါတယ်။၅ မိနစ် စမ်းကြည့်
check_permission ကို non-destructive tool တွေကိုလည်း optional allowed_paths argument နဲ့ ကန့်သတ်နိုင်အောင် တိုးချဲ့ပါ (ဥပမာ read_file က request လုပ်ထားတဲ့ path သည် sandbox directory အတိအကျတစ်ခုအထဲမှာမှသာ အောင်မြင်သင့်ပါတယ်)။ sandbox အတွင်းက read_file request အောင်မြင်တာနဲ့ sandbox အပြင်က တစ်ခု deny ခံရတာကို ပြသတဲ့ test call တွေ ရေးပါ၊ read_file ကိုယ်တိုင် allowlist ထဲပါပြီး non-destructive ဖြစ်နေတောင်မှ။
သတိလေးတစ်ချက်
Destructive tool call တစ်ခုအတွက် 'confirmed' ဒါမှမဟုတ် 'approved' flag ကို LLM ရဲ့ output ကိုယ်တိုင် သတ်မှတ်ခွင့်ပြုခြင်း — model ကိုယ်တိုင် ၎င်းရဲ့ permission ကို grant လုပ်နိုင်ရင် confirmation step ဟာ real gate မဟုတ်တော့ဘဲ decorative ဖြစ်သွားပါတယ်။
'Flexible ဖြစ်အောင်' ဆိုပြီး agent တစ်ခုကို tool သီးသန့် narrow မလုပ်ဘဲ general-purpose shell tool ပေးခြင်း — open shell တစ်ခုက least-privilege ကိုလည်း allowlist ကိုလည်း တစ်ပြိုင်နက်တည်း ချေဖျက်ပစ်ပါတယ်။
Anthropic — Tool Use Overview — Local AI / Local LLM