Agents Scenarios

Model calls, context, tools, and coordinated agent execution.

All domains · Classes A–Z

Source locations

• Inputs and calculations: mlsysim/engine/scenario_definitions/agents.py.
• Named configurations: mlsysim/engine/scenario_catalog/agents/.
• Result schemas: mlsysim/engine/result_types/agents.py.

Public imports and existing class reference URLs remain stable.

Scenario contracts

• AcceptedTaskCostScenario: An attempt of model tokens plus other cost, retried until accepted; optional levers on it.
• AllocationAxesScenario: One output ceiling spent as one chain, candidates streams, or rounds of feedback.
• AnnFootprintScenario: Resident bytes of a graph index and of product-quantization codes over one corpus.
• AppendOnlyContextScenario: A trajectory that appends each turn to its prompt and resends the whole history.
• ApprovalEscrowScenario: Approval escrow for a long-running workload (held resources) and an urgent one (time).
• ArgmaxNoiseScenario: Argmax over equally good candidates with Gaussian score noise; a rare loophole in the pool.
• AttentionDilutionScenario: One attention head with fixed logits: a target among background keys, optionally look-alikes.
• BestOfNPoolScenario: Pools of independent candidates at pass_rate, and a mixed workload with the same average.
• BlindRetryScenario: Mutating calls retried once after any timeout, with independent request and reply loss.
• Bm25PairScenario: One query term scored by BM25 in a short and a long document of one corpus.
• BoundedLiabilityAdmissionScenario: The next model call of a trajectory, tested against its dollar allowance.
• CheckerPosteriorScenario: A fallible check over candidates, defect_rate of which violate the property.
• CohortArmRecord: One policy’s outcomes and settled usage over a cohort of attempted tasks.
• CompactionCadenceScenario: Compaction between a trigger and a target watermark over a turn budget.
• ContextAdmissionScenario: A serialized request (text plus an image charge) against a window with a reserved reply.
• ContextBudgetScenario: A request budgeted by role, appending each tool result against selecting a fixed share.
• ContextInvalidationScenario: A superseded excerpt right after the reused prefix: remove, replace, or append a notice.
• ControlledBranchScenario: Fresh continuations from one saved prefix under several branches, counted by outcome; the Wilson bound a clean branch leaves and the clean runs a target_bound needs.
• CorrelatedReviewersScenario: Reviewers who each falsely approve with the same marginal rate, independent or correlated.
• CounterfactualSearchScenario: Replay paths that blame for a failure must search: a whole horizon, or one checked window.
• CreditWindowScenario: A producer feeding a slower model consumer, unbounded or behind a credit window.
• DeadlinePolicyScenario: One candidate, breadth in parallel, or feedback, for one task under one deadline.
• DeficitRoundRobinScenario: Backlogged request classes sharing model calls under weighted deficit round-robin.
• DelegationCohortScenario: A single controller and a delegated design on one matched cohort, priced alike.
• DelegationTokenBreakEvenScenario: Delegation at a token multiple m per attempt: cost per accepted task scales by m p_single / p_multi.
• DependencyScheduleScenario: A dependency graph on interchangeable slots: aggregate work, the critical path, and both bounds.
• EnvelopeAmortizationScenario: calls sequential calls of work_per_call, with envelope setup paid per call or once.
• ExactScanScenario: An exact vector scan on one host: bytes, intensity, and the bandwidth floor.
• FanoutDeadlineScenario: A fan-out whose workers each finish by the deadline with one probability, independently.
• GraphNeighborhoodScenario: A typed traversal’s exact-tree neighborhood against a call’s retrieval allowance.
• HeadSamplingMissScenario: Head sampling keeps each trial with retention_probability, decided before its outcome; a failure mode that occurs occurrences times leaves no detailed trace with miss.
• HorizonCompoundingScenario: Independent, unchecked steps each right with step_success; success at each horizon.
• IndexMaintenanceScenario: One edit’s synchronous re-indexing against exclusion of the stale entries.
• InjectionExposureScenario: Workloads reading reads untrusted items each, a fraction payload_fraction carrying a payload, with and without a screen that misses miss_rate of payloads.
• ModelCallScenario: One call of prompt_tokens (+ instruction_tokens) in and output_tokens out at floors.
• ModelPollingScenario: A model that checks a pending job at a fixed interval until it completes.
• ObservationCarryScenario: A raw tool output sized in windows, and its shaped view carried by later calls.
• PartitionScreenScenario: A fixed partition screened over worker counts: (1 - p) + p / N + c (N - 1) of the serial time.
• PolicyComparisonScenario: Two policies, each run on attacked attacked and benign separate benign episodes.
• PrefixLayoutBillScenario: A trajectory of root, trunk, and leaf requests, billed volatile-first and zoned.
• ProbeVsAnswerScenario: A generated explanation of answer_tokens against a prepared test of probe_time.
• ProtectedFieldCheckScenario: Constraints lost independently across compaction rounds, with and without a host check.
• PrunedSearchTreeScenario: A search tree of branching children per node whose step check rejects rejection_rate.
• ReferencePassingScenario: Children handed a pasted repository slice, or a reference and the files they then read.
• ResidualViolationScenario: A residual per-step rate of violating proposals over horizon steps, for a day’s trajectories.
• RetrievalPayloadScenario: Whole serialized records staged in what a call’s window leaves for retrieval.
• SamplingScenario: One fixed logit vector at a base and a sharpened temperature, then top-p at the base.
• SandboxPoolReserveScenario: A warm sandbox pool under steady arrivals: active leases, a standby reserve, and memory.
• ScalabilityLawScenario: Relative capacity N / (1 + sigma (N - 1) + kappa N (N - 1)) over integer worker counts.
• ScopedRevisionScenario: A candidate plan of equal nodes, replaced_nodes of which a refuted premise invalidates.
• SearchAccountsScenario: Flat best-of-N against beam search over the same task, kept as separate accounts.
• SequentialCanaryScenario: A canary monitor: Wald’s SPRT between acceptance p_base and a specified degraded p_degraded at targets alpha (false hold) and beta (missed degradation), on fraction of eligible tasks arriving at arrival_rate, labels label_delay late.
• SerialTailExposureScenario: Serial calls that each exceed their own tail latency independently.
• SharedWriteScenario: Workers each writing writes_per_worker of files files, uniformly and independently.
• ShortenedWindowScenario: An observation window cut to window_share of its adopted length; samples scale with it.
• SpanCriticalPathScenario: One trial’s observed spans on a shared clock, the spans each waits for, and one intervention.
• StorageAmplificationScenario: Auxiliary bytes and re-index time of a lexical-plus-graph and a lexical-plus-dense store.
• StreamCancelScenario: A streamed reply with output_allowance tokens, rejected and canceled at reject_step.
• SyscallTaxScenario: calls ordinary system calls priced by the terms each isolation envelope pays.
• ToolCatalogSelectionScenario: A registered tool catalog staged whole or as a selected subset with overhead.
• ToolDriftScenario: Baseline and candidate distributions of tool calls over the same tools, compared by the Jensen-Shannon divergence (natural logarithms).
• ToolkitGranularityScenario: Structured tools against one shell tool, judged by cost per accepted task.
• TraceRetentionScenario: One cohort period of trials in disjoint verdict classes, each with a trace_size body, retained over window by keeping everything, head sampling, or verdict-aware selection.
• TrajectoryBudgetMeterScenario: An append-only trajectory metered against turns, tokens, time, and dollars.
• TrajectoryResendScenario: A trajectory of turns calls that each resend the whole history (TrajectoryResendModel).
• VerifiedPoolScenario: A pool of pool_size candidates, first passing one returned, under a fallible check.
• WorkspaceProvisioningScenario: workspaces on one pinned base, provisioned by copy or copy-on-write, then a first write.