{"results":[{"id":"quickstart","title":"Quickstart","url":"/docs/quickstart","content":"Version, trace, replay, and prove your first AI agent release: create an Agent Genome, run with signed tracing, verify the Evidence Ledger, generate an Evidence Package, and connect Agenomic over MCP. In about 10 minutes you will create an Agent Genome, run an agent with signed tracing, verify the signed Evidence Ledger, generate an Evidence Package, evaluate technical evidence for EU AI Act workflows, and optionally connect the MCP Server so agents can write their own evidence trail. Some commands below show the target Agenomic CLI surface; where one is not yet wired in your build, treat it as planned and substitute your internal equivalent. Install the `agenomic` CLI with Homebrew, the shell installer, or npm, then confirm the version. If a package channel is not published for your environment yet, treat that command as planned and replace it with your internal install command. Every local step below runs offline, with no account. An Agent Genome is the signed, portable record of what defines an agent release: prompts, models, tools, policies, permissions, dependencies, memory contracts, traces, metrics, replay reports, and evidence. Scaffold one with `agenomic init`. A minimal genome pins the model, system prompt, tools, policy governance rules, and memory contract. Resolve and validate the genome before you run it. Verification confirms that tools, policies, and dependencies resolve, and records the canonical genome hash. Agenomic captures a canonical trace for each run. Events are append-only, hash-chained, and can be sealed into a tamper-evident Evidence Ledger. Each run emits a canonical, append-only sequence of events. Agenomic does not require storing raw prompts or tool payloads in logs — production setups store hashes and redacted payloads by default. Verification proves the technical integrity of the recorded run: the hash chain, the Merkle root, and the signature. It does not by itself provide legal certification. An Evidence Package collects the genome snapshot, signed Evidence Ledger, verification report, metrics, and compliance mapping into one signed archive. Verify a package end to end before you share it. Never present a simulation or synthetic replay as probative evidence — Agenomic labels simulations as non-probative. Agenomic maps runs and releases to technical evidence workflows for risk management, technical documentation, record-keeping, and human oversight. It generates technical evidence and supports EU AI Act workflows; it does not certify compliance, and legal compliance requires qualified review. Agenomic compares the Agent Genome and the observed behavior, not just source files, so reviewers see behavioral drift before promotion. Replay re-derives a run for verification. `exact` mode attempts a hash match; when a provider is non-deterministic, Agenomic falls back to functional equivalence and reports a fidelity score. Evidentiary replay proves a run from signed artifacts with no re-execution — the basis for audit-ready evidence. Agenomic supports five replay modes, from strict reproduction to proof without re-execution. The SDK mirrors the CLI for programmatic runs, events, verification, and evidence. Keep secrets and raw payloads out of trace bodies. Agents can use Agenomic directly through MCP. The Agenomic MCP Server lets compatible LLMs, agents, IDEs, and runtimes call Agenomic tools while they run. Add it to your client config — this is a first-class path, not an afterthought. The server exposes tools for runs, genomes, policy governance, evidence, compliance, metrics, drift, and replay. A typical agent governs its own actions over MCP: start a run, propose a tool call, check policy, request human review when required, execute, complete, verify, and emit evidence. MCP outputs are redacted by default — raw prompts, completions, tokens, secrets, and sensitive payloads are not exposed unless explicitly enabled in a controlled environment. Before Agenomic handles production traffic, turn on authentication, isolation, signing, and observability. A representative production configuration. Agenomic should emit structured JSON logs, OpenTelemetry traces, Prometheus metrics, and request and trace IDs, plus worker job, Evidence Ledger verification, Evidence Package, compliance evaluation, replay, and drift metrics. By default, Agenomic logs and exposes integrity metadata, never raw sensitive content. You have gone from a local run to a signed Evidence Package and connected agents to Agenomic over MCP. Dig into the Agent Genome model in Concepts, the CLI reference and Python SDK, and the EU AI Act evidence guides linked below; revisit the MCP Server, Evidence Package, replay, and drift sections above for production detail. To start for real, create your first genome with `agenomic init`."},{"id":"cli-reference","title":"CLI reference","url":"/docs/cli-reference","content":"Core `agenomic` commands for local release, replay, diff, and cloud publishing. `agenomic run` executes a workflow definition, captures trace events, and writes an agent.lock file for repeatable comparison. When connected to Agenomic Cloud, the CLI uploads signed bundles to the registry for review and replay."},{"id":"python-sdk","title":"Python SDK","url":"/docs/python-sdk","content":"Use Python to submit runs, attach metadata, and query release evidence. The SDK mirrors the cloud API while keeping secrets outside trace payloads. Evidence records should avoid real personal data and include enough metadata for audit review."},{"id":"cloud-guide","title":"Cloud guide","url":"/docs/cloud-guide","content":"Connect teams, registries, approvals, and AI Act evidence packs. Cloud registries organize bundles by workspace and expose approval workflows for regulated releases. Reviewers compare signed attestations, replay summaries, and behavior contract drift before promotion."},{"id":"concepts","title":"Concepts","url":"/docs/concepts","content":"Agents, workflows, runs, bundles, locks, and replay distributions. An agent is the behavior boundary you release. A workflow is the reproducible execution plan used to produce run evidence. A run captures traces and outputs. Replay compares distributions and contract outcomes instead of pretending LLM behavior is perfectly deterministic."},{"id":"eu-ai-act-compliance","title":"EU AI Act compliance for AI agents","url":"/docs/eu-ai-act-compliance","content":"How Agenomic produces signed, replayable evidence that maps to EU AI Act obligations (Articles 9–15) for production AI agents. Agenomic provides evidence; it does not issue legal certification. The EU AI Act regulates AI systems by risk tier. Agentic systems used in high-risk contexts inherit obligations that are largely about evidence: you must be able to show how the system was built, how it behaves, and that you keep records over its lifetime. Most of these obligations live in Articles 9–15. Agenomic encodes each agent into a signed genome and builds a reproducible bundle from it. That genome, its execution traces, and its attestations are the raw material for AI Act evidence — captured as part of your release flow rather than reconstructed by hand before an audit. Agenomic is strongest on the documentation, record-keeping, and behavioral-evidence obligations. It supports — but does not replace — your own risk-management and oversight processes. Agenomic produces evidence that maps to the AI Act; it does not certify compliance, guarantee legal conformity, or constitute legal advice. Conformity assessment, classifying your system’s risk tier, and sign-off remain your responsibility. Agenomic’s job is to make the evidence portable, signed, and easy to produce. Capture a genome and a trace with the quickstart, then read the two evidence guides below. When you are ready to share evidence for review, Agenomic Cloud can collect signed bundles into AI Act evidence packs with approval workflows — entirely optional, since the CLI and SDKs run offline."},{"id":"ai-act-technical-documentation","title":"EU AI Act technical documentation for AI agents (Article 11 & Annex IV)","url":"/docs/ai-act-technical-documentation","content":"Article 11 and Annex IV of the EU AI Act require technical documentation drawn up before a system reaches the market and kept up to date. Here is how an Agenomic genome and bundle capture much of it automatically. Article 11 requires high-risk AI systems to have technical documentation, and Annex IV lists what it must contain. Much of it is a structured description of the system: what it is, how it was built, and how it behaves. An Agenomic agent-bundle is a structured, versioned snapshot of an agent. The files in a bundle line up with several Annex IV elements, so the documentation tracks the system instead of drifting from it. Because the bundle is built and signed per release, your technical documentation is reproducible and diffable rather than a document that goes stale. Agenomic captures and structures evidence that maps to Annex IV; it does not write your full technical documentation for you, classify your system’s risk tier, or certify compliance. Treat the genome and bundle as the backbone you attach narrative and assessments to."},{"id":"agent-audit-trail","title":"Build an audit trail for AI agents (EU AI Act Article 12)","url":"/docs/agent-audit-trail","content":"Article 12 of the EU AI Act requires automatic record-keeping over a system’s lifetime. Here is how Agenomic’s signed traces, attestations, and replay give AI agents a tamper-evident audit trail. Article 12 asks high-risk AI systems to automatically record events (“logs”) over their lifetime, so the system functioning is traceable and post-incident review is possible. For an AI agent, the events that matter are the ones that explain why it did what it did. An agent’s behavior is more than its final answer. Agenomic captures the execution trace — the sequence of model calls, tool calls, inputs, and outputs — alongside the genome that defines the agent, so a record stays meaningful months later. Capturing logs is not enough if they can be edited after the fact. Agenomic signs genomes and emits tamper-evident attestations (ATEP), so an archived record can be verified rather than trusted. Traces export to JSONL for your own storage and pipelines. A static log shows what happened once. Replay re-runs an agent and compares output distributions and behavior-contract outcomes against a baseline, so your audit trail also shows whether behavior held over time. Replay is statistical — it does not assert bit-for-bit identical results. Follow the quickstart to capture your first trace, then archive the signed bundle from each release. Agenomic provides this evidence; it does not issue legal certification, and Article 12 sign-off remains your responsibility."},{"id":"piste-audit-agents-ia","title":"Créer une piste d’audit pour les agents IA (article 12 du règlement IA)","url":"/docs/piste-audit-agents-ia","content":"L’article 12 du règlement européen sur l’IA impose une journalisation automatique tout au long du cycle de vie du système. Voici comment les traces signées, les attestations et le replay d’Agenomic dotent les agents IA d’une piste d’audit infalsifiable. L’article 12 du règlement sur l’IA demande aux systèmes à haut risque d’enregistrer automatiquement des évènements (« logs ») tout au long de leur cycle de vie, afin que le fonctionnement du système soit traçable et qu’une analyse post-incident soit possible. Pour un agent IA, les évènements qui comptent sont ceux qui expliquent pourquoi il a agi ainsi. Le comportement d’un agent ne se résume pas à sa réponse finale. Agenomic capture la trace d’exécution — la séquence d’appels au modèle, d’appels d’outils, d’entrées et de sorties — aux côtés du génome qui définit l’agent, pour qu’un enregistrement reste exploitable des mois plus tard. Enregistrer des logs ne suffit pas s’ils peuvent être modifiés après coup. Agenomic signe les génomes et émet des attestations infalsifiables (ATEP), de sorte qu’un enregistrement archivé peut être vérifié plutôt que cru sur parole. Les traces s’exportent en JSONL pour votre propre stockage. Un journal statique montre ce qui s’est passé une fois. Le replay réexécute un agent et compare les distributions de sorties et les résultats du contrat de comportement à une référence : la piste d’audit montre aussi si le comportement s’est maintenu dans le temps. Le replay est statistique — il n’affirme pas des résultats strictement identiques. Suivez le quickstart pour capturer votre première trace, puis archivez le bundle signé de chaque version. Agenomic fournit ces preuves ; il ne délivre pas de certification légale, et la validation au titre de l’article 12 reste de votre responsabilité."}]}