Sheaft

Resilience intelligence for networked and agent systems

Sheaft turns telemetry, execution traces, topology and domain constraints into a live reliability model, stress-tests disruption scenarios, and shows which critical operations or agent workflows are fragile before failures propagate.

Built for graph-structured systems: software infrastructure, agent harnesses, smart mobility, and scientific reaction or interaction networks.

View GitHub See use cases

One resilience model for graph-structured and agent systems

Sheaft represents a system as an attributed graph: nodes, edges, constraints, telemetry, execution traces, stressors and critical operations. The domain changes; the resilience question stays the same: what breaks, how far it propagates, and which operations or workflows are affected.

Nodes

Services, agent steps, tools, road segments, sensors, proteins, metabolites, reactants.

Edges

Calls, tool invocations, state reads/writes, flows, roads, reactions, interactions.

Attributes

Latency, capacity, confidence, permission state, retrieval age, energy, rate, state.

Constraints

Domain contracts, policy checks, evaluator gates, signal rules, conservation laws, biological priors.

Stressors

Crashes, tool timeouts, schema drift, stale retrieval, missing observations, perturbations.

Verdict

Fragile paths, blast radius, affected operations or workflows, green/yellow/red posture.

How the MB3R toolchain works

Procrustes checks whether the available static evidence is sufficient. Bering discovers the runtime graph. Sheaft evaluates disruption scenarios and produces resilience verdicts.

  1. Check model eligibility

    Procrustes checks deployment, tracing and reproducible-operation artifacts and prepares a conservative collection plan for Bering.

  2. Discover the system graph

    Bering ingests telemetry, topology files, execution traces, event streams or explicit graph descriptions and builds a typed system model.

  3. Attach domain contracts

    Critical operations, constraints and success predicates are attached to the graph: service journeys, agent workflows, corridor operations, reaction routes or biological pathways.

  4. Stress-test disruption scenarios

    Sheaft simulates failures, missing observations, degraded components, tool timeouts, schema drift, stale retrieval, demand spikes or perturbations without breaking the real system.

  5. Produce a resilience verdict

    The result is a graph-level report: fragile components, affected operations, blast radius, posture trend and recommended validation targets.

Six high-value domains. One resilience engine.

Most mature: public MVP + experiments

SRE and digital infrastructure

Sheaft evaluates resilience of software systems before release and between releases. It builds service graphs from trace data or topology artifacts, simulates dependency failures, and returns a gate or posture verdict for critical user journeys.

Digital infrastructure is graph-structured operational infrastructure. Services, APIs, queues, databases, replicas, release policies and user operations form one dependency graph. Validation evidence

Inputs

  • Traces/OTLP
  • Service topology
  • Endpoint contracts
  • CI/CD artifacts
  • Incident history

Outputs

  • Release-risk verdict
  • Fragile dependencies
  • Affected endpoints
  • Chaos-test priorities
  • Posture history
Engineering research line

Agent reliability and AI harnesses

Sheaft builds a reliability model from agent harness execution traces: LLM calls, tools, retrieval, memory, policy checks, evaluators, permissions, retries, fallback paths and human gates.

Single-run task success and average pass rate miss boundary failures. Real incidents appear between the model, tools, retrieval, state, guardrails and release gates. OpenTelemetry GenAI conventionsAgent spans

Inputs

  • Execution traces / GenAI spans
  • Tool calls
  • Retrieval and memory reads/writes
  • Policy/evaluator/human-gate events
  • Retry/fallback history

Outputs

  • Reliability graph
  • Virtual chaos scenarios
  • Fragile boundaries
  • Green/yellow/red release verdict
  • Test priorities
Pilot-ready

Smart mobility and traffic corridors

Sheaft evaluates resilience of smart mobility infrastructure represented as a live attributed graph: road segments, intersections, toll gates, sensors, signal controllers, payment systems and operations centers.

Smart mobility is cyber-physical infrastructure. Road networks, tolling, sensors and digital services form one distributed operational graph. NIST CPS contextSUMO road graph

Inputs

  • Road topology
  • Traffic/toll events
  • Sensor observations
  • Signal timings
  • Incident logs

Outputs

  • Fragile corridors
  • Missing observations
  • Incident propagation
  • Affected mobility journeys
  • Operator resilience verdict
Research and partner pilots

Chemistry and biology networks

Sheaft analyzes chemical reaction networks, biological interaction networks, metabolic pathways and bioprocess graphs to find fragile routes, critical intermediates, observation gaps and perturbation-sensitive modules.

These systems are routinely modeled as graphs: reactions connect reactants, intermediates and products; biological networks connect proteins, genes, metabolites and signaling interactions. EMBL-EBI biological networksRSC reaction networks

Inputs

  • Reaction networks
  • Pathway models
  • Biological interaction graphs
  • Experimental observations

Outputs

  • Critical intermediates
  • Alternative routes
  • Fragile modules
  • Perturbation blast radius
Pilot-ready

Warehouse logistics

Sheaft evaluates inbound, storage, replenishment, picking, packing and shipping as one operational graph and tests how congestion, equipment outages or demand spikes affect order fulfilment.

Locations, routes, buffers, equipment, inventory and order waves form a changing dependency network rather than a set of isolated warehouse KPIs. GS1 EPCIS visibility events

Inputs

  • Warehouse topology and routes
  • WMS and EPCIS events
  • Inventory and buffer states
  • Equipment telemetry
  • Order waves and SLAs

Outputs

  • Bottlenecks and fragile routes
  • Order and SLA risk
  • Congestion propagation
  • Recovery paths
  • Equipment and staffing priorities
Partner pilots

Production planning and scheduling

Sheaft represents production plans as a graph of jobs, operations, materials, equipment, personnel and precedence constraints, then evaluates plan robustness under delays, shortages, changeovers and asset downtime.

A feasible plan can still be fragile: one late material, unavailable machine or quality hold may propagate across orders and delivery commitments. ISA-95 operations modelNIST digital twins

Inputs

  • ERP/MES plans and job routes
  • Materials and BOM dependencies
  • Equipment calendars and capacity
  • Personnel and shift constraints
  • Execution and quality events

Outputs

  • Schedule robustness
  • Critical resources and constraints
  • Delay propagation
  • Affected orders and commitments
  • Recovery and replanning priorities

Three stages: eligibility, model discovery and resilience assessment

Procrustes: static eligibility preflight

Procrustes checks whether deployment, tracing and reproducible-operation artifacts provide enough static evidence for Bering to attempt an operation-specific, replica-aware runtime model.

  • Resolves services, images, identities, entrypoints and replica counts.
  • Checks tracing prerequisites without collecting telemetry.
  • Matches a target HTTP operation to a reproducible scenario.
  • Emits findings, remediation, a collection plan and a conservative Bering overlay.
Procrustes on GitHub Procrustes architecture

Bering: system-graph discovery and artifact publishing

Bering builds typed graph artifacts from telemetry, topology inputs, execution traces, event streams or explicit domain models. It publishes stable model and snapshot artifacts for downstream resilience analysis.

  • Discovers typed graph structure from traces, OTLP, GenAI spans, event streams or explicit topology.
  • Normalizes nodes, edges, attributes, state reads/writes and domain contracts.
  • Publishes stable model/snapshot artifacts.
  • Keeps provenance so teams can inspect where each part of the model came from.
Bering on GitHub Bering docs

Sheaft: resilience simulation, verdicts and posture history

Sheaft consumes graph artifacts, stress-tests disruption scenarios, evaluates domain policies and tracks resilience posture or release verdicts over time.

  • Runs simulation-based resilience analysis on graph artifacts.
  • Evaluates policy rules for critical operations and agent workflows.
  • Reports fragile components, affected operations and blast radius.
  • Supports batch checks and continuous posture tracking.
Sheaft on GitHub Sheaft docs

Validated on distributed systems, extending to agent harnesses and broader infrastructure

ICSE 2026 Distinguished Paper Award

Model Discovery and Graph Simulation was recognized in ICSE 2026 NIER with a Distinguished Paper Award.

Official ICSE page

DeathStarBench experiments

Graph-discovered resilience models were evaluated against live fault-injection outcomes on a distributed benchmark.

OpenTelemetry GenAI/agent semantics

OpenTelemetry is forming GenAI, agent and tool-span semantics, which makes trace-derived reliability models a portable research abstraction rather than a local log format.

GenAI conventions Agent spans

Public MVP

Bering and Sheaft are available as public tools for graph discovery, simulation, verdicts and posture monitoring.

GitHub

What this proves

Sheaft can turn passive telemetry and execution traces into an explicit model and produce useful resilience signals without running broad live experiments every time.

Demo report: digital infrastructure use case

The current public report shows the most mature software-infrastructure workflow. The same report pattern extends to agent harnesses, mobility and scientific-network pilots.

Smart mobility report template

Corridor graph, missing observations, affected mobility journeys, disruption propagation.

Chemistry/biology report template

Reaction/pathway graph, critical intermediates, perturbation-sensitive modules, alternative routes.

Agent harness report template

Trace-derived call graph, stale retrieval, tool timeout, schema drift, permission denial, release verdict.

Warehouse logistics report template

Flow graph, bottlenecks, congestion propagation, at-risk orders, recovery paths.

Production-plan report template

Job graph, critical constraints, delay propagation, affected commitments, replanning priorities.

Start with one graph, one critical operation, one historical period

SRE pilot

Start from trace data, topology and incident history for one service domain. Sheaft builds a model, runs failure scenarios, and compares the result with known incidents or release risks.

Agent reliability research pilot

Start from one agent workflow and one set of historical execution traces. Sheaft builds the harness reliability graph, simulates virtual chaos scenarios, and returns a green/yellow/red release verdict.

Smart mobility pilot

Use topology and telemetry from one corridor, tolling domain, parking/payment flow or sensorized mobility area. Sheaft identifies fragile components, missing observations and affected mobility journeys.

Chemistry/biology research pilot

Use a reaction network, biological interaction graph or pathway model with perturbation scenarios. Sheaft identifies critical intermediates, fragile modules and robustness hypotheses.

Warehouse logistics pilot

Start with one warehouse zone, one order flow and a representative historical period. Sheaft identifies bottlenecks, propagation paths and recovery options against actual fulfilment outcomes.

Production-plan pilot

Start with one production plan, its resource and material constraints, and execution history. Sheaft stress-tests delays and outages and ranks the commitments and replanning actions at risk.

Research base: graph models, resilience and emergence

Sheaft is grounded in model discovery, graph simulation, sheaf-theoretic consistency, causal emergence and resilience monitoring. The product packages these ideas into practical workflows for networked and agent systems.

Research extensions

  • Chemistry/biology network resilience
  • Traffic corridor resilience
  • Agent harness reliability and virtual chaos simulation
  • Warehouse-flow resilience
  • Production-plan robustness

Presentation