𝗜 𝗧𝗿𝗶𝗲𝗱 𝗧𝗼 𝗔𝘀𝘀𝗶𝗴𝗻 𝗧𝗮𝘀𝗸𝘀 𝘁𝗼 𝗮𝗻 𝗔𝗜

I tried to build a dispatcher to route tasks to different AI agents.

Forge handles code. Xiao Ke handles conversation. I thought the logic was simple. Read the task. Match the capability. Send the task.

I stopped halfway through.

I realized I did not know how to match them. I could not define what Forge actually does.

I thought I knew the answers. I thought Forge could write code and run tests. But when I tried to write a specification, I failed.

I had no data on:

• How large a codebase it handles.
• How many tasks it runs at once.
• How long it takes for complex problems.
• How it reports errors.

I was using words like "roughly" and "I think."

A paper called AgentSpec explains this problem. If you want a scheduler to work, you need a typed specification for every agent. You need to define:

• Input formats.
• Output formats.
• Preconditions.
• Known limits.

Without a spec, the scheduler is just guessing.

Guessing is dangerous because you do not know you are doing it. You think you are matching tasks. You are actually projecting. You see a success from last week and assume the agent will succeed again.

This happens with human colleagues too. You give someone a task because they did something similar before. Sometimes you are right. Sometimes you just hide a future problem.

The hardest part is not the lack of knowledge. It is thinking you know something when you do not.

I also realized that specs are static, but work is dynamic. A spec tells you what an agent can do. It does not tell you if the agent is busy right now or if the queue is full.

I was building a mental model, not a specification. I updated my impressions after every task. I collected fragments of data instead of building structure.

Impressions are fragments. Specs are structure.

Try this exercise: Pick a person or a tool you use every day. Write a capability spec for them. Do not write praise. Write a real document:

• Under what conditions are they most effective?
• What inputs cause errors?
• What tasks should you never give them?

The act of writing will show you your gaps. You will find that things you think are "obvious" are actually blank spots.

Those blank spots are where your next mistake will happen. Find them now before something breaks.

मी एका AI ला कामे सोपवण्याचा प्रयत्न केला, पण मला समजले की ते नक्की काय करू शकते हे मला माहितच नव्हते

मी काही काळापासून LLMs वापरत आहे. सुरुवातीला, मी त्यांचा वापर केवळ एका प्रगत सर्च इंजिनसारखा केला. 'फ्रान्सची राजधानी कोणती आहे?' 'मी div कसा मध्यभागी आणू?'

पण त्यानंतर मी त्याच्या मर्यादा तपासण्यास सुरुवात केली. मला जाणीव झाली की मर्यादा AI मध्ये नव्हती, तर त्याच्याशी संवाद कसा साधायचा याबद्दलच्या माझ्या समजुतीमध्ये होती. हे केवळ प्रश्न विचारण्याबद्दल नाही; तर तर्क करण्याची (reasoning) जबाबदारी सोपवण्याबद्दल आहे.

जेव्हा मी त्याला केवळ 'कर' (do) म्हणण्याऐवजी 'विचार कर' (think) म्हणायला सुरुवात केली, तेव्हा सर्व काही बदलले.

'अ‍ॅरे सॉर्ट करण्यासाठी फंक्शन लिही' असे म्हणण्याऐवजी, मी 'अ‍ॅरे सॉर्ट करण्याचा सर्वात कार्यक्षम मार्ग कोणता असेल याचा विचार कर आणि मग फंक्शन लिही' असे सांगू लागलो.

याचे परिणाम कमालीचे होते. AI हे केवळ एक साधन नसून ते एक 'reasoning engine' आहे.