Agentic Loops¶
Az Agentic Loop az a runtime execution model, amelyben a rendszer egy goal eléréséhez ismételten observe → decide → act → observe ciklust futtat, miközben canonical state, budget, policy és stop condition deterministic control alatt marad.
Hasznos mental model:
Goal
↓
Observe
↓
Decide
↓
Act
↓
Observe result
↓
Continue / Replan / Ask / Stop
A loop nem egyszerűen „LLM gondolkodik újra és újra”. A model egy decision component; a runtime tulajdonolja az execution lifecycle-t.
Roadmap¶
1. Agentic Loop Mental Model¶
Agentic loop vs egyszerű model call; workflow vs loop; deterministic shell és bounded autonomy.
Status: first detailed draft complete.
2. Goals, Tasks and Success Conditions¶
User requestből explicit execution goal, scope, constraint, subtask és verifiable success condition.
Status: first detailed draft complete.
3. Observation and Environment State¶
Authoritative current state, observation, freshness, fact vs hypothesis, normalized tool result és environment feedback.
Status: first detailed draft complete.
4. Decision and Action Selection¶
Structured next-action contract, tool/skill selection, ASK/WAIT/REPLAN/STOP, deterministic vs model-driven transition.
Status: first detailed draft complete.
5. Planning and Task Decomposition¶
Plan-first, reactive és short-horizon planning, task decomposition, dependency graph, replanning és plan fixation.
Status: first detailed draft complete.
6. Execution State and Scratchpad¶
Canonical run state, context projection, scratchpad, artifact, checkpoint, resume és concurrency/versioning.
Status: first detailed draft complete.
7. Stop Conditions and Execution Budgets¶
Success/failure termination, iteration/tool/time/token/cost budget, no-progress detection, cancellation és nested-loop budget.
Status: first detailed draft complete.
8. Failure Recovery, Retry and Replanning¶
Failure classification, retry vs repair vs re-observe vs replan, idempotency, reconciliation és graceful degradation.
Status: first detailed draft complete.
9. Reflection, Verification and Critique¶
Self-reflection vs independent verification, deterministic verifier, critic pattern és generate → verify → repair.
Status: first detailed draft complete.
10. Human-in-the-Loop and Approval Gates¶
Clarification, approval, escalation, exact-action approval scope, expiry, revalidation és durable pause/resume.
Status: first detailed draft complete.
11. Loop Patterns and Control Strategies¶
Observe/act, plan/execute, planner/executor, verification loop, deterministic workflow + agentic island, supervisor/worker és event-driven patternök.
Status: first detailed draft complete.
12. Loop Evaluation, Observability and Anti-patterns¶
Outcome + trajectory evaluation, trace, cost/latency/tool efficiency, replay/failure injection és loop anti-patternök.
Status: first detailed draft complete.
Tanulási sorrend¶
- 1. Agentic Loop Mental Model
- 2. Goals, Tasks and Success Conditions
- 3. Observation and Environment State
- 4. Decision and Action Selection
- 5. Planning and Task Decomposition
- 6. Execution State and Scratchpad
- 7. Stop Conditions and Execution Budgets
- 8. Failure Recovery, Retry and Replanning
- 9. Reflection, Verification and Critique
- 10. Human-in-the-Loop and Approval Gates
- 11. Loop Patterns and Control Strategies
- 12. Loop Evaluation, Observability and Anti-patterns
Core architecture stance¶
canonical state
↓
purpose-specific context
↓
model proposes next action
↓
validation / authorization / budget
↓
capability execution
↓
new observation
↓
state transition
A model proposal és a real side effect között mindig legyen explicit runtime boundary.
Készültségi állapot¶
Az Agentic Loops workstream first-draft mélységben elkészült. A következő terület az Agent Architecture, amely a skilleket és loopokat teljes software systemként helyezi el.