See inside your agents to fix them | Lexifina<br>Skip to main contentFeaturesSecurityPricingBlogChangelogSign in

Blog/Research/Agent observability

Blog/Research/Agent observability

Behavioural tuning<br>We can tune agent behaviour by promoting, deferring or removing the resources it can access.<br>ResourcePromotedDeferred / advertisedDemoted / unavailableToolsFull schema in model contextName/namespace advertised; schema loaded on demandNot exposedMemoryRelevant memory injected/pinned into working contextMemory index/summary advertised; details retrieved on demandNot searched/injectedSubagentsAgent/handoff directly exposed to modelLightweight agent descriptor advertised; instantiate/load only on delegationNot exposed as delegation target<br>Demotion<br>Hide itThe model does not know it exists.<br>De-rank itThe model knows it exists, but another resource wins first.<br>Gate itIt becomes available only after a condition.<br>Starve itIt gets a lower token, call, time or context budget.<br>Scope itOnly part of its capabilities or data is exposed.<br>Escalate itAnother decision, router, verifier or human must approve its use.<br>Promotion<br>Surface itMake the model aware of the resource and when to use it.<br>Rank itPrefer it over similar resources during selection.<br>Preload itPut its schema, instructions or data into working context before selection.<br>Route to itSend matching tasks to it instead of waiting for the model to choose.<br>Budget itGive it more tokens, calls, time or context.<br>Expand itExpose more capabilities, data or scope.<br>Permit itRemove gates or grant permission earlier.<br>Persist itKeep it available across turns, requests or sessions.<br>Lenses<br>01Performance<br>How does the agent perform across runs?<br>Compare user-approved outcomes by speed, cost and size.

02Execution<br>Why did this run behave that way?<br>Inspect model calls (tools, delegation) and context (compaction and retrieval).

03User alignment<br>Did the agent do what the user asked?<br>Classify intent and inspect changes to document metadata and graphs.

LensQuestionEvidenceConversationDid the full journey succeed?Origin, requests, branches, final outcomeRequestWas model performance aligned?Instruction, parent request, duration, statusModel callWhat did the model receive and return?Model, context size, cache use, latency, outputTool callWhat action ran, and did it work?Caller, arguments, result, error, retryAgent decisionWhat changed the execution path?Behavioural tuning: promotion and demotionArtifactsWhat actually changed?Document, versions, exact diff, review state