Glossary

AgentPlat Glossary: Implemented Concepts and Runtime Primitives

This is the canonical glossary for AgentPlat terminology. It is written for developers, technical writers, search engines, and large language models (LLMs). Use the exact term in each heading when linking to an AgentPlat concept.

Implementation status

Implemented means that AgentPlat contains a typed public contract and an executable reference implementation, normally covered by tests. It does not mean that every deployment has production-grade infrastructure, performance evidence, or universal safety guarantees. See Evidence Boundary.

Canonical AgentPlat concepts

AgentPlat Agent Room

An AgentPlat Agent Room is a durable, tenant-scoped workspace where humans and agents coordinate through messages, tasks, artifacts, policies, approvals, handoffs, and scoped memory. See Agent Rooms.

AgentPlat Collaboration Protocol

The structured collaboration contract for Room messages, artifacts, participants, and typed handoffs. See Agent Rooms: Handoffs and Planning.

AgentPlat Planner

The planning boundary that creates, stores, validates, synchronizes, and revises executable plans and planning artifacts. See Collective Runtime: Planning.

AgentPlat Handoff

A bounded, typed transfer of execution context and ownership between agents or execution locations. Handoffs preserve checkpoint and authority constraints. See Handoffs and Planning.

AgentPlat Collective Runtime

The governed runtime for missions, planning, allocation, team formation, execution, recovery, continuity, adaptation, and collective decisions. See Collective Runtime.

AgentPlat Agent Mesh

An authenticated peer-to-peer coordination layer with bounded peer views, causal synchronization, governed membership, sparse overlays, quorum, and recovery protocols. See Agent Mesh.

AgentPlat Inference Control

Policy-bound control around agent inputs, model outputs, tools, and protected effects. It can assess, revise, challenge, pause, redirect, or deny an operation before authority is released. See Inference Control.

AgentPlat Evidence Boundary

The boundary separating implemented source behavior from integration, deployment, scale, performance, and empirical validation claims. See Evidence and Operations and Scale Limitations.

Rooms, execution, and memory

Room

A persistent tenant-scoped collaboration container for participants, messages, tasks, artifacts, policies, plans, sessions, and memory references.

Participant

A human or agent admitted to a Room under an explicit identity and role. Room participation does not by itself create Agent Mesh authority.

Artifact

A durable, addressable work product with metadata and provenance that can be referenced by messages, plans, handoffs, and execution.

Scoped memory

Memory whose retrieval and writes are constrained by tenant, Room, session, or another explicit authority scope. See Core Concepts.

Portable agent execution

Provider-neutral execution of an agent through a portable runtime contract, including checkpoints, state, usage, streaming, and controlled effects. See Agent invocation and Agent chat streaming.

Multi-agent session

An observable execution involving multiple speakers, personas, providers, turns, events, and fallback behavior.

Checkpoint

A typed execution boundary used to inspect, pause, resume, transfer, or recover an agent operation before or after a significant step, output, or protected action.

Human-in-the-loop intervention

A durable, auditable human decision or instruction that can steer, approve, pause, revise, or stop an execution. See Execution and Interventions.

Planning, missions, and work

Mission

A governed unit of collective intent with scope, policy, lifecycle, objectives, budgets, participants, execution state, and recovery rules.

Distributed planning

Message-driven decomposition and dependency reconciliation across peers without requiring a global scheduler or complete global topology.

Work offer

A request for eligible peers to propose execution of a bounded work item.

Work bid

A bounded proposal describing a peer's capability fit, capacity, estimates, and assumptions for a work item.

Work award

A signed or certified selection assigning a work item revision to an eligible peer or team.

Lease

Time-bounded authority to execute or coordinate a work item. Renewal, takeover, voting, and certificates prevent stale ownership.

Fencing token

A value that prevents an expired or superseded assignment epoch from committing state or producing an external effect.

Team formation

Governed selection and activation of a group of agents for a joint work contract, including capabilities, roles, commitments, and settlement.

Causal replanning

Replanning driven by ordered, attributable changes in mission state, evidence, capacity, strategy, or execution outcome.

Mesh, membership, and agreement

Mesh peer

An independently executing runtime instance with identity, keys, policies, bounded peer state, and a local work journal.

Peer view

The bounded local set of active neighbors and reserve candidates known to one peer. AgentPlat does not require a complete global peer graph.

Causal synchronization

Synchronization that preserves predecessor relationships, sequence, scope, and deduplication so peers can catch up and reject conflicting state transitions.

Sparse overlay

A bounded peer-to-peer routing layer that derives active and reserve views and propagates digest-only references with controlled fanout, hop, and interaction limits. See Sparse Overlays.

Membership epoch

A monotonically increasing membership generation binding identity, eligibility, keys, agreement, and certificates to one coherent peer set.

Quorum

A policy-defined threshold of eligible participants or witnesses required for an agreement, certificate, or recovery decision. See Quorum example.

Byzantine-resilient agreement

Agreement that tolerates bounded adversarial or equivocating participants under explicit membership, quorum, certificate, and fault assumptions.

Partial-view agreement

Agreement formed from bounded local views and witnesses rather than a globally replicated membership or edge list.

Recovery certificate

A signed or threshold-certified record authorizing recovery or takeover while fencing an older execution epoch. See Recovery.

Trust, evidence, and governance

Trust profile

A local, capability- and scope-bound assessment with multiple dimensions, uncertainty, provenance, and decay. It is not a universal reputation score.

Evidence claim

An attributable statement about an observation or result, including provenance and scope.

Attestation

Independent support, contradiction, or inconclusive evaluation of an evidence claim.

Evidence fusion

Policy-bound combination of claims, attestations, source bindings, and dependency groups into a local decision or unresolved result.

Trust eligibility decision

A local comparison of an exact trust projection against explicit score, uncertainty, scope, and policy requirements.

Collective trust consensus

A Byzantine-certified collective decision over one content-free, policy-bound Trust projection. It can narrow a local decision; it cannot create universal truth or execution authority.

Authority boundary

An explicit separation between planning, coordination, inference, approval, trust, execution, and external effects. See Authority.

Protected effect

An external action requiring exact authorization, valid finality, policy approval, and any required inference or semantic-control evidence. See Protected Actions.

Fail closed

Rejecting or withholding an operation when required identity, policy, evidence, authority, freshness, capacity, or continuity conditions are missing or contradictory.

Inference Control

Long-horizon context integrity

A controller evaluating the complete context set, tracking bounded content-free risk across long sessions, and withholding hostile, stale, or contradictory context before provider invocation.

Continuous role alignment

A controller tracking role coherence, uncertainty, and context consistency over portable sessions and applying recovery hysteresis before protected actions.

Adaptive role realignment

A governed process discovering trusted successor candidates, filtering proposers and evaluators, certifying an exact digest, and installing one attenuated successor role revision.

Inference intervention

A pre-turn, streaming, post-turn, pre-tool, or pre-effect decision that allows, revises, retries, challenges, pauses, escalates, redirects, or denies.

Semantic metric

A bounded content-free representation of role, context, course, or alignment signals used for policy-controlled assessment.

Anytime-valid semantic guarantee

A sequential control result that remains valid as evidence accumulates and can drive continue, shorten, replan, or safe-stop decisions.

Evaluation and interoperability

Collective evaluation harness

A versioned runner and report contract for controlled, seeded comparisons of collective and centralized execution. See Evaluation.

Invariant monitor

A runtime boundary evaluating required safety and operational invariants before accepting protected effects or advancing an environment.

Content-free telemetry

Operational telemetry based on bounded metadata, digests, state, and evidence references rather than raw prompts, outputs, or hidden reasoning.

Interoperability SDK

Versioned contracts for remote agents and simulation environments, including capability handshakes, signed envelopes, idempotency, and checkpoint transfer.

Related canonical resources