A Transaction Science Platform

The agent runtime substrate.
MCP + A2A native.

Plan / Step / Action / Observation. Tool sandbox. Five-tier memory. Multi-agent orchestration. Jailbreak guardrails. OTel-shape traces. Trajectory eval harness. Provider adapters via inv-agent-adapters. Every agent step emits a signed JWP receipt.

API health ↑ Transaction Science

Capabilities

Nine crates. The agent loop, end to end.

AGENT LOOP

Plan / Step / Action / Observation / Decision

The reduced ReAct + planner-executor loop every modern agent runs. Strongly-typed events; AgentEventKind for Started / Planned / ActionTaken / Observed / Replanned / Halted / HumanRequested / Completed. Receipt-stamped at each transition.

MCP

Model Context Protocol native

Server capabilities + tools + resources + prompts. JSON-RPC envelope. Open standard for tool, resource, and prompt exposure.

A2A

Agent-to-Agent protocol

Agent Cards at /.well-known/agent.json. Capabilities + skills + auth-scheme negotiation. Task lifecycle (Submitted → Working → InputRequired → Completed).

TOOLS

Sandboxed tool registry

Per-tool side-effect classification (None / ReadOnly / LocalWrite / NetworkRead / NetworkWrite / ExternalEffect / Destructive). Permission grants at registry level — FsRead/Write/NetEgress/Process/Secrets — checked before every call.

MEMORY

Short / Long / Episodic / Semantic / Procedural

Five-tier memory abstraction with ring-buffer eviction for short-term, full-store for long-term. Source-receipt traceability per entry.

MULTI-AGENT

Sequential / Parallel / Hierarchical / PlannerWorker / Debate / RouterDispatch

Six orchestration patterns. Dispatch plans with inter-step depends_on.

GUARDRAILS

Input + output filters

Jailbreak / prompt-injection / PII / secret / toxicity / hallucination / harmful / out-of-scope / data-exfiltration detection. Verdict: Allow / Redact / Block / HumanReview.

TRACES

OpenTelemetry-shape spans

Run / Plan / LlmCall / ToolCall / Retrieval / Subagent / Decision span kinds. Parent-child tracking. Each span emits a JWP receipt.

EVALS

Trajectory grading

OutputContains / OutputJsonAt / ToolCalled / NoToolCalled / NoErrors / StepsBelow / JudgePasses criteria. Cost-budget enforcement. Pass/fail with criterion-level diagnostics.

Framework adapters

Provider-adapter surface.

Claude Agent SDK

inv-agent-adapters::claude_sdk

MCP-native; Python + TypeScript

OpenAI Agents SDK

inv-agent-adapters::openai_agents

Handoffs + guardrails first-class

LangGraph

inv-agent-adapters::langgraph

State-machine model with checkpointing

AWS Strands Agents

inv-agent-adapters::strands

AWS-native agent runtime

SDK

Pull what you need.

[dependencies]
inv-agent          = "0.1"
inv-tools          = "0.1"
inv-memory         = "0.1"
inv-multiagent     = "0.1"
inv-a2a            = "0.1"
inv-guardrails     = "0.1"
inv-trace          = "0.1"
inv-eval-agent     = "0.1"
inv-agent-adapters = "0.1"

Thesis

Receipt-stamped agents are court-admissible.

Every agent step — Plan / Action / Observation / Decision — and every tool call inside it emits a JWP ReceiptPayload. For regulated finance, healthcare, and gov customers, that's the difference between an agent decision the auditor can verify and one they can't.

Pairs with EngineeringOS (author-time) and DeliveryOS (pipeline-time). Three pillars, one signed receipt format.

Request AgentOS access

The agent runtime substrate. MCP + A2A native.