Enterprise Generative Search — Azure and Open Source  ·  View 23 of 41  ·  Runtime

The Bounded Agentic Loop

Where the model is allowed to decide, and the three things it never decides.

Editable source SVG draw.io All views
Understand class and entities Plan sub-queries, backends Retrieve policy-filtered Evaluate evidence coverage and conflict Refine or stop budget decides Synthesise claims, not prose Verify grounding and safety Bounded Loop 3 hops, 18 s typed plan retrieval requests candidates and reasons gap named, not felt one more hop, or enough draft with claim markers answer, abstain or fall back The Bounded Agentic Loop — Where the Model Is Allowed to Decide Application we own Data store Decision point Security / platform The model chooses what to ask for. It never chooses what it may see, how long it may run, or whether its answer ships. v 1.0 · owner Data and AI Global Practice

The decision

  • The model chooses what to ask for. It does not choose what it may see, how long it may run, or whether its answer ships. Those three are held by the entitlement resolver, the budget governor and the verifier respectively.
  • The loop is a state machine with typed transitions, not a free-running agent. Every transition is a span in the trace, which is what makes a wrong answer diagnosable rather than merely regrettable.
  • Evaluate evidence names the gap — coverage of a sub-query, or a detected conflict — so refinement is targeted retrieval rather than another attempt at the same question.

Numbers

  • Hard limits: 3 hops, 45,000 tokens, 18 seconds, 6 retrieval calls. Reaching any limit produces the best grounded answer so far, explicitly marked as incomplete.
  • Measured: 2.4 hops average on the multi-hop class; 7% of agentic queries hit a limit, and their task-success rate is tracked separately.

Risks

  • An agentic loop with no ceiling is a cost incident waiting for one badly phrased question. The budget governor is a hard stop, not a warning.
  • Loops that fail late are expensive. The evaluate stage is deliberately cheap so that stopping early is the common outcome.
◀  Ask to Answer — The Critical Path All views Hybrid Retrieval and Fusion  ▶
Solution Architecture v1.0 · Data & AI Global Practice · 2026-09 · 41 views · Microsoft Azure and open source23-agentic-loop