Manifold Grace

Proposed educational extension · 4 November 2026

Agentic Architecture: From Automated Workflows to Verifiable Parallel Intelligence

AI agents can combine planning, memory, tools and autonomous task selection with deterministic business-process automation. The resulting systems vary from a single agent with tools to layered, distributed, event-driven and cloud–edge multi-agent networks. No one topology is inherently AGI or ASI: exceptional generalisation and reproducible performance must be established independently.

Source and illustration policy: These are original explanatory schematics informed by architectural categories in industry guidance and the supplied diagram references. They are not copies of vendor figures, deployed systems, safety certifications or evidence that an ASI system exists. Original proposal records remain on their own dated pages.

Illustration A · Twelve-direction research coordination

Authorised requestIntent, scope, permissions
OrchestratorDependency graph, resource budget
01Evidence
02History
03Definitions
04Contradictions
05Causality
06Counterfactuals
07Stochastic models
08Forecasts
09Procedures
10Cross-agent state
11Risk
12Outcomes
Evidence-aware reconcilerUnresolved claims retained
Independent assessorTest reasoning, rights, safety
Decision or escalationHuman approval for sensitive actions
Fan-out/fan-in is conceptual: parallel work is valid only for independent tasks. Downstream checks depend on earlier evidence and may require retries, cancellation or human intervention.

Ten topology patterns and trade-offs

TopologyControl flowSuitable applicationsKey failure mode
Single agent and toolsOne scoped planner with retrieval, memory and task-specific toolsLow-complexity, bounded workloadsUnnecessary tool permissions; unbounded execution
Sequential pipelineAgent A → Agent B → verifierDocument processing, staged transformationsError propagation; serial latency
Supervisor–workerPlanner → specialist workers → reconcilerCross-domain evidence reviewCentral bottleneck; correlated source assumptions
Hierarchical teamsCoordinator → domain managers → specialist agentsMany interdependent domainsComplexity, tracing and compounded costs
Concurrent fan-out/fan-inIndependent branches → evidence-aware mergeTwelve-direction checks, independent analysesConflicting results, data races, high compute
Handoff / triageClassifier or expert → appropriate specialistUnexpected issues requiring specialist expertiseRouting loops, lost context and wrong delegation
Iterative maker–checkerProposer ⇄ separate evaluator, with capped retriesResearch synthesis and document validationSelf-confirmation, costly nonconvergence
Peer-to-peer meshAuthenticated agent-to-agent exchangesFederated cross-organisation discoveryUnclear ownership, message loops and trust boundaries
Event-driven BPA / agent mixScheduled/webhook events → deterministic rules → scoped agentNews digest, claims triage and back-office processingDuplicate events, stale context, unsafe actions
Cloud–edge–deviceCloud coordination ↔ edge inference ↔ telemetryIndustrial monitoring and latency-sensitive systemsNetwork partitions, unsafe actuation and version mismatch

Choose the simplest architecture that reliably meets the task, rather than multiplying agents by default. Hybrid workflows can switch between deterministic sequential steps, parallel specialist checks and human-reviewed interventions as dependencies and risk require.

Illustration B · Hybrid orchestration and control planes

Events / scheduled jobsWebhooks, documents, telemetry
Deterministic workflow engineValidation, deduplication, retries
Agent reasoning planePlanner, specialists, memory, retrieval
Policy and approval gateRead-only / write / high impact
Scoped API executionMCP-backed tools, RPA or services
Independent monitoringAudit, outcomes, rollback
Hybrid automation separates probabilistic recommendations from deterministic transaction execution. A model's request to call a tool is not proof that its action is authorised.

Illustration C · Cloud, edge and device deployment

Cloud planeAnalysis, model registry, retraining, audit
Edge planeLocal inference, offline fallback, safety boundaries
Device / sensor planeTelemetry and constrained actuation
Physical control cannot rely solely on an unverified text model. Existing certified safety controls, locally enforced limits, and human operational responsibility must remain separate.

Operational and assurance requirements

Identity and authority

Separate user identity, agent identity and delegated permissions; least-privilege tool scopes; time-bounded tokens.

MCP and tool calls

Tool discovery and invocation are interfaces, not permission to act; validate arguments and outputs, enforce policy outside the model.

Agent-to-agent exchanges

Typed schemas, versioned contracts, sender authentication, message integrity, replay protection and trace IDs.

Memory and knowledge

Distinguish context-window state, durable task state, cited knowledge, inferred conclusions and update histories.

Concurrency correctness

Idempotency keys, ordered event logs, locks/transactions, deadline propagation, cancellation and bounded queues.

Safety gates

Default read-only operation; explicitly approved writes; simulation before actuation; escalation and abort paths.

Observability

Trace agent paths, model versions, retrieved sources, tool invocations, latency, cost and corrections.

Reliability

Retries with backoff, circuit breakers, timeouts, checkpoints, disaster recovery and manual fallback.

Verification

Independent source checks, calibration, causal validity, task-level rubrics and expert adjudication.

Privacy and rights

Purpose limitation, minimisation, retention controls, meaningful contestability and restricted data access.

Four worked architecture selections

1. Conflicting AI infographic analysis — supervised parallel verification

Extract graph and text claims with source positions; send independent date, definition, forecast and contradiction checks; reconcile claim IDs and evidence; preserve uncertain or ambiguous relationships. No external writes are required.

2. Institutional casework — hierarchical and procedural assurance

Process records through dated rules, accessibility checks, deadlines, procedural standards and evidence registries. Specialist assistants may recommend action; a designated person reviews rights-sensitive decisions, with reasons and challenge routes recorded.

3. Research/news digest — event-driven automation and bounded synthesis

A scheduled workflow collects multiple feeds, de-duplicates records, validates timestamps and sources, summarises distinct claims, retains original links and routes a proposed digest for review. Event retries should be idempotent. A failed agent should not trigger repeated public posting.

4. Industrial monitoring — cloud/edge event hierarchy

Sensors provide telemetry to an edge gateway with deterministic alarm thresholds and network-partition safeguards. Cloud agents support diagnostics and planning. Physical interventions remain under independent operational safety and authorised human control.

Evaluation beyond final-answer accuracy

Compare a single-agent baseline, tool-enabled agent, multi-agent design without independent assurance, and the corresponding design with an independent verifier. Assess all four on equal-cost and equal-wall-clock budgets, reporting both rather than collapsing them into a single score. Use held-out novel tasks, grounded claims, calibrated uncertainty, human expert adjudication, causal validity, consistency, idempotent execution, latency, token and API spend, permissions, reversibility and error recovery.

Record exact model and tool versions, dataset dates, seeds, timeout policies, retries and approval interventions. Parallel agents sharing one underlying model are not necessarily independent evidential sources. More complexity can worsen performance and costs.

Research integration

Read alongside the Verifiable Advanced Intelligence framework, AI Research Lab, Evidence Verification Across Sectors, Information Science and Feature Request and Contribution Register. This page introduces new educational illustrations; it does not alter the historical records of prior proposals.

Architecture references

Research addition dated 4 November 2026. Educational design study, not a deployed workflow. No third-party diagrams reproduced.