Home · Library · Graph

Topic Map

A visual guide to the relationships between production architecture, context, evaluation, governance and agent operations.

Topics and relationships

Agentic AI

Agentic AIMulti-Agent Topologies · Memory · Context Architecture · Reasoning · Governance

Systems in which a language model plans, calls tools, and acts over multiple steps rather than answering a single prompt. The demo takes a weekend and the platform takes a year, because orchestration, memory, evaluation, and governance all turn out to be load-bearing. Most of the machine side of this library is about making such systems survive contact with a regulated enterprise.

Multi-Agent Topologies · direct dependency
Once a single agent becomes several, the topology of who may call whom becomes the primary design decision.

Memory · direct dependency
An agent that acts over multiple steps needs some persistence of state, and that persistence is where much of the risk concentrates.

Context Architecture · direct dependency
Agent behaviour is largely determined by what context reaches the model at each decision point.

Reasoning · direct dependency
Planning and tool use in agents are built directly on the reasoning stack of chain-of-thought, search, and test-time compute.

Governance · application
Governance is what agentic AI has to pass through before it is allowed to act in a regulated enterprise.

Multi-Agent TopologiesAgentic AI · Guardrails

The wiring diagram of who talks to whom in a system of cooperating agents. Topology decides blast radius: treating tool permissions as a reachability graph lets you compute, before deployment, the maximum damage a single misbehaving agent could cause. The shape of the graph is a risk decision, not a framework default.

Agentic AI · direct dependency
Once a single agent becomes several, the topology of who may call whom becomes the primary design decision.

Guardrails · related topic
Blast radius analysis treats the permission graph and the communication graph as one object, so topology and guardrails are designed together.

ReasoningAgentic AI · Evaluation · Uncertainty

The stack of techniques that let models work through problems: chain-of-thought, tool use, search, reinforcement learning on verifiable rewards, and test-time compute. The layers vary widely in how well the evidence holds up. Spending more inference compute is now a deliberate lever with a real invoice, to be pulled against the value of the error it prevents.

Agentic AI · direct dependency
Planning and tool use in agents are built directly on the reasoning stack of chain-of-thought, search, and test-time compute.

Evaluation · application
Process reward models score the steps of a reasoning trajectory, not just its outcome, which is what trajectory assurance requires.

Uncertainty · related topic
Deciding how much test-time compute a step deserves is a judgement about how uncertain that step really is.

UncertaintyReasoning · Evaluation

Knowing how confident the system should be, and routing work accordingly. The certainty gradient places each workflow step at the cheapest tier its true ambiguity justifies, with calibrated cascades moving requests up only when confidence is genuinely low. Miscalibration, not raw error rate, is what makes automated decisions dangerous.

Reasoning · related topic
Deciding how much test-time compute a step deserves is a judgement about how uncertain that step really is.

Evaluation · shared problem
Both depend on calibration, since a confidence score you cannot trust is worse than no score at all.

Context Architecture

Context ArchitectureAgentic AI · Memory · Retrieval

The discipline of deciding what a model sees at each decision point. Context quality, not model quality, is what separates agentic systems that work from ones that flail. The strongest pattern is reconstruction over accumulation: rebuild context fresh from source-of-truth systems rather than letting it silt up indefinitely.

Agentic AI · direct dependency
Agent behaviour is largely determined by what context reaches the model at each decision point.

Memory · related topic
Reconstruction over accumulation is a memory policy expressed as a context policy: rebuild from source of truth instead of trusting what has silted up.

Retrieval · direct dependency
Context is assembled from what retrieval returns, so retrieval quality sets a ceiling on context quality.

MemoryAgentic AI · Context Architecture · Retrieval · Evaluation · Guardrails · Identity Resolution

What an agent carries forward between steps and sessions, from scratchpads to long-term case stores. Memory is not a passive record; it is an active input to future autonomous decisions, which makes every unchecked write an attack surface and every long horizon a drift risk. In the enterprise the durable memory layer is often a governed warehouse, not a vector database.

Agentic AI · direct dependency
An agent that acts over multiple steps needs some persistence of state, and that persistence is where much of the risk concentrates.

Context Architecture · related topic
Reconstruction over accumulation is a memory policy expressed as a context policy: rebuild from source of truth instead of trusting what has silted up.

Retrieval · direct dependency
Stored memory is only useful through the retrieval path that brings the right piece of it back at the right moment.

Evaluation · shared problem
Memory drift over long horizons is invisible to outcome metrics, so both fields need sentinels that watch behaviour rather than answers.

Guardrails · shared problem
A memory write is an untrusted boundary crossing, so provenance and quarantine on writes are guardrails applied inside the agent.

Identity Resolution · shared problem
A memory store is only as trustworthy as its answer to whose fact this is and which entity it belongs to.

RetrievalMemory · Context Architecture

How an agent finds the material its context is built from. Single-signal vector search fails in enterprise settings; production accuracy comes from hybrid retrieval that fuses dense, sparse, and structured signals and then reranks with a cross-encoder. Retrieval quality sets a hard ceiling on everything downstream.

Memory · direct dependency
Stored memory is only useful through the retrieval path that brings the right piece of it back at the right moment.

Context Architecture · direct dependency
Context is assembled from what retrieval returns, so retrieval quality sets a ceiling on context quality.

Governance

GovernanceAgentic AI · Evaluation · Guardrails

The evidence problem behind every deployment in a regulated environment. Getting an agent live in a bank is less about the technology than about producing artefacts a risk committee can sign off with understanding rather than trust. Autonomy ledgers, handoff design, and trajectory-level assurance are how that evidence gets made.

Agentic AI · application
Governance is what agentic AI has to pass through before it is allowed to act in a regulated enterprise.

Evaluation · direct dependency
The evaluation harness produces the evidence that a risk function needs before it will sign off a deployment.

Guardrails · related topic
Governance decides what an agent should be allowed to do and guardrails are the mechanisms that enforce the decision.

EvaluationMemory · Governance · Reasoning · Uncertainty

The machinery that tells you, automatically and before your users do, whether the system still works. Prompts are disposable and models are rented, so the eval harness of golden cases, adversarial scenarios, and CI gates is the durable asset. Behavioural drift shows up in tool-selection distributions long before accuracy metrics move.

Memory · shared problem
Memory drift over long horizons is invisible to outcome metrics, so both fields need sentinels that watch behaviour rather than answers.

Governance · direct dependency
The evaluation harness produces the evidence that a risk function needs before it will sign off a deployment.

Reasoning · application
Process reward models score the steps of a reasoning trajectory, not just its outcome, which is what trajectory assurance requires.

Uncertainty · shared problem
Both depend on calibration, since a confidence score you cannot trust is worse than no score at all.

GuardrailsMulti-Agent Topologies · Memory · Governance · Identity Resolution

The hard constraints around what an agent may do, enforced outside the model rather than requested inside the prompt. In practice this means scoped capability tokens minted per tool call, network perimeters like VPC Service Controls, and platform-level content filters. Guardrails are the precondition that lets a bank run agents at all.

Multi-Agent Topologies · related topic
Blast radius analysis treats the permission graph and the communication graph as one object, so topology and guardrails are designed together.

Memory · shared problem
A memory write is an untrusted boundary crossing, so provenance and quarantine on writes are guardrails applied inside the agent.

Governance · related topic
Governance decides what an agent should be allowed to do and guardrails are the mechanisms that enforce the decision.

Identity Resolution · related topic
Scoped capability tokens only work if the identity behind each tool call is resolved and attributable.

Identity ResolutionGuardrails · Memory

Working out who and what the records refer to: which customer, which agent, which fact belongs to which entity. For agents this splits in two, resolving entities in the data they act on, and pinning down the agent's own identity through workload identity and per-task credentials so that every action is attributable.

Guardrails · related topic
Scoped capability tokens only work if the identity behind each tool call is resolved and attributable.

Memory · shared problem
A memory store is only as trustworthy as its answer to whose fact this is and which entity it belongs to.

Connect: