Agent Infrastructure Landscape

An open, comparative catalog of the tools developers use to build, deploy, and operate autonomous AI agents — the software systems (Claude Code, Cursor, AutoGen, LangGraph, Mem0, Zep, etc.) that take instructions, plan, call tools, remember things across sessions, and act in the world.

Each entry is tracked across 85 attributes (license, maturity tier, deployment model, MCP / A2A protocol support, observability stack, compliance posture, latency, pricing, and dozens more), with typed relationships between systems (built-on, extends, competes-with, cites…). Today the catalog covers 912 systems and 528 relationships.

The goal is straightforward: give builders, researchers, and analysts a single place to compare what exists, see what's actually shipping vs. just published, and spot the gaps.

Start here

• Browse the catalog → https://mrpeppersdev.github.io/agent-infrastructure-landscape/ — sortable, filterable table with full per-system detail pages, side-by- side comparisons, lineages, leaderboards, and "best of" lists (e.g. open-source agent memory, MCP-enabled systems).
• Headline findings → five short, citable stories distilled from the catalog. Read them as cards on the site or in source: docs/FINDINGS.md.
• Raw data → data/landscape.json + data/landscape.edges.json. Licensed CC-BY-4.0 — use it however you like, just credit the source.

What's in the data

Each record is a JSON object with a stable ID, display name, tier (1 battle-tested → 5 theoretical), one or more section memberships (e.g. "Dedicated memory layers", "Framework-embedded memory", "Vector-database infrastructure" — 34 sections in total), and the 85 attribute cells. Every cell carries:

• The claim itself (free-text value).
• A source URL (where the claim came from).
• A status (real-data, estimate, not-applicable, no-data, etc.).
• A provenance tier (T1 auto-verifiable from a GitHub URL, T2 resolvable source URL required, T3 estimate).
• A last-verified date for high-volatility cells.

Sources include curated lists (Agent-Memory-Paper-List, Awesome-GraphMemory), survey papers, benchmark leaderboards (LongMemEval, LoCoMo, ConvoMem), vendor sites, academic venue pages, and targeted research-agent sweeps. Claims are vendor-stated unless otherwise marked. Honest coverage confidence is roughly 88–92% — known gaps are tracked in PLAN.md. Full schema: docs/SCHEMA.md.

Use it in your own tools

MCP server — query the catalog from any Model Context Protocol client (Claude Code, Claude Desktop, Continue, etc.). Nine read-only tools: search, get-by-id, edge traversal, coverage stats, side-by-side comparison, recent changes, eval-orphan detection, substrate blast-radius analysis. See mcp/ and mcp/README.md.

cd mcp && npm install && npm run build
claude mcp add landscape -- node $PWD/dist/server.js

CLI — the same nine queries as a landscape <subcommand> tool. Text output by default, --json / --csv for machine-readable use. See cli/ and cli/README.md.

cd cli && npm install && npm run build
./dist/landscape.js search "memory" --tier 1 --section "Dedicated memory layers"

Web app — SvelteKit static export in web/, deployed to GitHub Pages on every push to main.

cd web && npm install
npm run dev      # local server at http://localhost:5173/
npm run build    # static export

How the catalog stays honest

Three things keep the data trustworthy:

1. Per-cell provenance. Every claim points at the source URL it was sourced from, with a tier marking how verifiable the claim is. You can audit any entry by walking the citation.
2. Automated freshness checks. A weekly job (.github/workflows/staleness.yml) flags rows whose upstream repo has gone quiet beyond the freshness SLA defined in MAINTAINER.md §2. Stale rows surface in the live table with a visible badge.
3. Self-maintaining workflows. New systems submitted via the /submit form (or an intake-labelled GitHub Issue) are auto-researched into a draft PR (docs/INTAKE.md). Existing sections are periodically re-audited section-by-section (docs/AUDIT.md).

Editing the catalog (contributors)

data/landscape.json is the source of truth — the rendered landscape.html is a build artefact, never edited by hand.

1. Edit data/landscape.json directly, or let the intake-research / section-audit bots open a PR for you. PRs get a rendered-cell preview comment from .github/workflows/diff-preview.yml so reviewers see exactly what each touched row will look like.

2. Run make build locally — reconciles JSON, rebuilds edges and citation trajectories, re-renders landscape.html.

3. Run make validate — five offline gates, ~25 seconds:

   #	Gate	Catches
   1	JSON schema	Records or edges violating docs/SCHEMA.md §7.
   2	Fast-step determinism	Non-determinism in extract / reconcile / edge-build scripts.
   3	Render-cycle stability	Markup drift in render ↔ extract round-trips beyond the documented ceiling.
   4	S2 cache integrity	Corrupted Semantic Scholar cache files.
   5	Tier-provenance + freshness	Cells whose tier disagrees with their citation, or rows missing required date metadata.

4. Commit the JSON, the regenerated edges file, and the regenerated landscape.html together. CI re-runs validation plus a byte-identity check (the JSON must round-trip cleanly to the committed HTML) on every push.

A pre-commit hook is available via make install-hooks (idempotent; short-circuits on changes that don't touch the pipeline). make refresh-citations re-pulls Semantic Scholar data (~15 min); only needed when adding research-paper rows.

Governance and license

This catalog has an explicit maintenance contract: MAINTAINER.md. It defines what's in scope (and what isn't), the freshness SLA per cell type, the 3-tier claim-validation schema, the succession plan, and how to contribute or request co-maintainer rights. The contract exists because comparative catalogs historically die quietly — DB-Engines, State of JS, dbdb.io — and the surviving ones all publish one.

The catalog data is released under CC-BY-4.0. The MCP server and CLI packages are MIT (see their package.json).

Credits

This catalog cross-references, corroborates, and (where licensing allows) imports data from the following external landscape projects. Vendored snapshots and the full per-source licence audit live in extraction/external/.

• Agentic Community Landscape (Apache-2.0) — CNCF landscape2 project cataloguing frameworks, protocols, memory, RAG, observability, and infra for agentic AI. Accessed 2026-07-01.
• SylphAI YC Agent Landscape (MIT) — AI-classified Y Combinator company dataset covering Winter 2024 – Fall 2026 batches with agent-vs-non-agent and platform-vs-single-purpose taxonomies. Accessed 2026-07-01.
• AI System Design Guide — Tool-Use and Computer Agents chapter (MIT) — narrative chapter on 2026 tool-use and computer-agent systems, cited for benchmark trajectories and adoption metrics. Accessed 2026-07-01.
• InclusionAI Agentic AI Landscape (no licence declared) — noted as prior art for landscape structure and category framing. Not imported; see extraction/external/LICENCES.md for the licence audit.