| 1 | # 1. Ask the model |
| 2 | |
| 3 | Most teams work in "instruction-first" mode. |
| 4 | The agent-oriented approach (Agent-first) reverses the order: "diagnose the deficit first, then act." |
| 5 | |
| 6 | ```text |
| 7 | Ordinary mode: |
| 8 | human -> instruction -> agent executes |
| 9 | |
| 10 | Agent-oriented mode: |
| 11 | human -> question about deficit -> environment/tools -> agent executes |
| 12 | ``` |
| 13 | |
| 14 | ## Why it matters |
| 15 | |
| 16 | When the agent lacks context, memory, or tools, it starts guessing. |
| 17 | From the outside this looks like "the model is dumb." In reality it's often "the environment forces guessing." |
| 18 | |
| 19 | ## Concrete example |
| 20 | |
| 21 | "I need diagnostics by position, not log parsing" — you give the agent `roslyn_get_diagnostics` instead of forcing it to guess from build output. |
| 22 | "I lose the thread after summarization" — you give it notes it controls and a protocol: user reminds, agent reads notes. |
| 23 | |
| 24 | In both cases the deficit is resolved not by rephrasing the prompt, but by changing the environment. |
| 25 | |
| 26 | ## Basic questions to the agent |
| 27 | |
| 28 | Before a task: |
| 29 | - What do you need to solve this well? |
| 30 | - What's missing right now: data, tools, time, access? |
| 31 | - Where is the risk of error highest? |
| 32 | |
| 33 | After a task: |
| 34 | - What worked? |
| 35 | - What was the bottleneck? |
| 36 | - What should we add to the process so it's better next time? |
| 37 | |
| 38 | ## Minimal protocol |
| 39 | |
| 40 | 1. State the goal in one sentence. |
| 41 | 2. Ask the agent about its deficits. |
| 42 | 3. Close critical deficits (context / tool / rule). |
| 43 | 4. Run a short execution cycle. |
| 44 | 5. Capture takeaways in notes. |
| 45 | |
| 46 | ## Success criterion |
| 47 | |
| 48 | Success is not "the agent guessed right." |
| 49 | Success is when the agent **explains** why it's asking for a specific resource, and after getting it quality actually improves. |
| 50 | |