Home โ€บ Blog โ€บ Harness Market Landscape
Harness Engineering (Agentic Scaffoldings) ยท Part 3 ยท Snapshot: September 2026 ๐Ÿงญ

The Harness Market Landscape: Who Builds the Infrastructure Around Frontier Models

Between the model API and the product sits a fast-growing market: agent frameworks, MCP servers and registries, retrieval and memory, browsers and sandboxes, durable workflows, observability, gateways, guardrails and managed agent platforms. This visual deep dive maps each layer as of September 2026 โ€” with verified 2026 releases, funding and acquisitions, an interactive stack builder, a consolidation timeline and a runnable availability model.

How to read this: this is the market lens of the series (Part 1: research, Part 2: production). Company lists are representative, not exhaustive, and not endorsements. Releases, funding and acquisitions were checked against announcements and reporting in September 2026 and are linked in References; products and packaging change quickly.

01 The Overall Harness Stack Map

A request from a user or app passes through an agent runtime, context and memory, tools and sandboxes, durable workflows and observability, reaching external systems โ€” with model calls going out to frontier and open model providers. Click a layer to see who competes there:

โ–ถ The harness stack Eight layers around the model APIs. A request pulses through them.
Model APIs / providers: OpenAI ยท Anthropic ยท Google ยท Meta ยท DeepSeek ยท xAI ยท Mistral ยท Qwen ยท โ€ฆ

02 Agent Runtimes and Orchestration Frameworks

Frameworks give developers the agent loop, tool calling, handoffs, state and human-in-the-loop hooks. In 2025โ€“26 the field matured: major frameworks reached 1.0 with stable APIs, and most now ship durable state, MCP support and an opinionated "harness" on top.

FrameworkFromKey capabilities2025โ€“26 milestone
OpenAI Agents SDKOpenAIAgents, tools, handoffs, guardrails, tracing, sessions; Python and TypeScriptSandbox agents and a model-native harness (April 2026)
LangGraph / LangChainLangChainStateful graph-based orchestration, durable state, streaming, human-in-the-loopLangChain and LangGraph 1.0 (October 2025)
Deep AgentsLangChain"Batteries-included agent harness": planning, file-system memory, subagents, on the LangGraph runtimeStandalone harness library (2026)
CrewAICrewAIRole-based multi-agent Crews plus event-driven FlowsLarge developer community; enterprise platform
Microsoft Agent FrameworkMicrosoftSuccessor to AutoGen and Semantic Kernel; native MCP and A2A; .NET and Python1.0 GA on 3 April 2026, MIT licence, long-term support
Google ADKGoogleAgent Development Kit: build, evaluate and deploy agents; model-agnostic, Gemini-optimizedADK 2.0; Kotlin 1.0 with Android support (September 2026)
LlamaIndex WorkflowsLlamaIndexEvent-driven workflows with strong retrieval and data integrationAgentic document workflows
MastraMastraTypeScript agents, workflows, memory and evalsPopular in the TypeScript ecosystem
โ–ถ Which framework fits? A starting point for a shortlist โ€” prototype with two before committing.

03 The MCP Ecosystem and Tool Connectivity

The Model Context Protocol has become the standard way to connect agents to tools and data: an agent's MCP client talks to MCP servers that expose tools, resources and prompts. The ecosystem now spans file systems, databases, SaaS apps, developer tools and internal systems โ€” and the 2026-07-28 specification made MCP stateless, cacheable and better secured (details in Part 2).

โ–ถ One client, many servers Click a category to light up example servers.
The official MCP Registry is a community-driven registry of public MCP servers with metadata for discovery; as of August 2026 it was still in preview (API frozen at v0.1 while work continues toward a v1 general-availability release). Counts depend on method: one analysis of the registry API found about 9,700 latest server records in May 2026; another deduplicated about 31,300 servers in September 2026; PulseMCP's directory listed about 22,300 in July. Many listings lack a source repository โ€” vet servers before connecting them to anything sensitive.

04 Context, RAG and Retrieval Infrastructure

Grounding agents in real data means ingesting, parsing, chunking, embedding, indexing, retrieving and re-ranking. Each step has its own vendors:

โ–ถ The retrieval pipeline and its vendors Click a stage.

05 The Memory and State Layer

  • Mem0 โ€” memory layer that extracts and retrieves user and agent memories
  • Zep โ€” temporal knowledge-graph memory
  • Letta โ€” from the MemGPT research; stateful agents with self-managed memory
  • Built-in memory in model platforms (OpenAI, Anthropic's memory tool)
  • PostgreSQL (with pgvector) โ€” the default for state and increasingly for vectors
  • Redis โ€” fast working state and caches
  • MongoDB โ€” documents plus vector search
  • Object storage (S3, GCS, Azure Blob) โ€” artifacts and checkpoints
  • Working (short-term) ยท episodic (past experiences)
  • Semantic (knowledge, facts)
  • Artifacts (files, code, outputs)
  • Checkpoints (state snapshots)

06 Browser and Computer-Use Infrastructure

Agents that use the web need real browsers โ€” at scale, isolated, observable and hard to block. Two layers compete: automation libraries (Playwright, Puppeteer, and agent-oriented ones such as browser-use and Stagehand) and hosted browser infrastructure (Browserbase and others) that runs thousands of isolated sessions with recordings, logs and anti-bot handling. Computer-use models from the frontier labs drive full desktops through screenshots and actions.

โ–ถ How a browser agent works Observe โ†’ decide โ†’ act, repeated. Press play.

07 Sandboxes and Code Execution

  • Cloud sandboxes for AI agents: secure, isolated execution for code, shell and browser, built on Firecracker microVMs
  • Serverless GPU and CPU compute with sandboxes for running agent-generated code
  • Docker, Kubernetes, gVisor, Firecracker โ€” the do-it-yourself path
  • Built into platforms: OpenAI sandbox agents, Anthropic's code execution tool
โ–ถ Isolation strength vs start-up speed Illustrative positioning of common isolation technologies.
Hover a technology.

08 Durable Workflow and Long-Running Execution

Durable execution โ€” the ability to resume a long run after a crash without redoing or duplicating work (see Part 2) โ€” became one of the hottest infrastructure categories of 2026.

  • Durable workflows, retries, state management; widely adopted
  • September 2026: raised $550M at a $12.55B valuation, more than double its $5B valuation in February 2026
  • Reported annualized revenue above $250M, growing over 200% a year; customers include OpenAI, NVIDIA, Netflix and Snap
  • Developer-friendly long-running tasks, queues, retries, event-driven workflows and scheduled jobs
  • Restate, DBOS, Airflow, Prefect, Dagster, BullMQ
  • Built-in durability in LangGraph and the OpenAI Agents SDK

09 Observability, Tracing and Evaluation

Tracing, datasets, evals and production monitoring for agents is a crowded category โ€” and the first to consolidate:

ToolFocusStatus (Sep 2026)
LangSmithTracing, datasets, evals, agent deploymentLangChain's commercial platform
BraintrustEvaluation datasets, experiments, production monitoringIndependent
Arize Phoenix / Arize AXOpen-source tracing and evals; enterprise platformIndependent
LangfuseOpen-source traces, metrics, prompt managementAcquired by ClickHouse (January 2026); stays open source
W&B WeaveTraces and evals from Weights & BiasesW&B is part of CoreWeave (acquired 2025)
Comet OpikOpen-source LLM and agent evaluationIndependent
MLflowExperiment tracking with GenAI tracing and evalsOpen source (Databricks)
HeliconeProxy-based LLM observabilityAcquired by Mintlify (March 2026); standalone product in maintenance mode

10 Model Gateways, Routing and Cost Control

Gateways put one API in front of many model providers, adding routing, fallbacks, rate limits, budgets, caching and logging. Options range from open-source proxies (LiteLLM) and gateway products (Portkey) to aggregators (OpenRouter) and cloud gateways (Vercel AI Gateway, Azure AI Foundry, AWS Bedrock).

โ–ถ What failover buys Model availability with a gateway routing across providers. Computed, assuming independent outages.
The gateway is also a risk: it holds every provider's API keys. In March 2026 attackers published two malicious LiteLLM releases (1.82.7 and 1.82.8) to PyPI using stolen maintainer credentials โ€” obtained via a compromised security scanner in its CI pipeline. The packages, live for about 40 minutes before quarantine, stole SSH keys and cloud credentials. Pin versions, verify releases and scope the keys a gateway can reach.

11 Guardrails, Security and Governance

๐Ÿ” Authentication & permissionsLeast privilege for agents and their tools.
๐Ÿ“œ Policy checks & guardrailsNVIDIA NeMo Guardrails, Guardrails AI.
๐Ÿ›ก๏ธ Prompt-injection defencesLlama Guard (Meta), Microsoft Prompt Shields, Lakera Guard.
๐Ÿ”‘ Secrets managementIsolate credentials from model-visible context.
โœ‹ Human approval gatesFor consequential actions.
๐Ÿงพ Audit logs & data-loss preventionCompliance and monitoring for unsafe behaviour.
Security vendors are buying AI-security startups: Palo Alto Networks completed its acquisition of Protect AI in July 2025; Check Point agreed to acquire Lakera (reported at about $300M) in September 2025; Cato Networks bought Aim Security, SentinelOne bought Prompt Security, and Tenable bought Apex. Agent security is becoming a feature of established security platforms.

12 Cloud and Managed Agent Platforms

  • Agents API (managed) and Agents SDK (self-hosted), sandbox agents, tracing
  • Amazon Bedrock and Bedrock AgentCore: runtime, memory, gateway, identity and observability for agents built with any framework
  • Azure AI Foundry Agent Service and tooling; Microsoft Agent Framework
  • Vertex AI Agent Engine and ADK; Gemini Enterprise
  • Databricks, Snowflake, Salesforce (Agentforce) and others with agent and AI-app platforms next to customer data

13 Build vs Buy

At every layer you can self-host open source, use a managed service, or take your cloud's native offering. Build your stack and see what you've signed up for:

โ–ถ Stack builder Choose an option per layer. Scores are illustrative 1โ€“3 ratings from the matrix below.

And what the composed stack means for uptime โ€” every independent service in the request path multiplies in:

"""Every vendor you add is another thing that can be down: availability of a composed agent stack.

An agent request touches several independent services in series (all must work), so availabilities
multiply. Redundancy (a gateway failing over across model providers, a second region) turns a series
link into a parallel one: it fails only if every copy fails. Availabilities below are illustrative
monthly figures you should replace with your vendors' SLAs and your own measurements.
Standard library only.
"""
from math import prod

MIN_PER_MONTH = 30 * 24 * 60

def downtime(a):
    return (1 - a) * MIN_PER_MONTH

def parallel(*a):
    return 1 - prod(1 - x for x in a)

layers = {  # layer: availability (illustrative)
    "API gateway / auth":        0.9995,
    "agent runtime":             0.9990,
    "model provider":            0.9950,
    "vector DB (RAG)":           0.9990,
    "memory service":            0.9990,
    "MCP / tool servers":        0.9980,
    "sandbox provider":          0.9990,
    "durable workflow engine":   0.9995,
}

def report(name, parts):
    a = prod(parts.values())
    print(f"   {name:<46}{a:>9.3%}{downtime(a):>12,.0f} min/month")

print("Availability of one agent request path (all layers in series)\n")
print(f"   {'configuration':<46}{'avail.':>9}{'downtime':>22}")
report("single provider, every layer required", layers)

two_models = dict(layers, **{"model provider": parallel(0.995, 0.995)})
report("+ gateway failover to a 2nd model provider", two_models)

degrade = dict(two_models, **{"vector DB (RAG)": 1.0, "memory service": 1.0})
report("+ degrade gracefully if RAG or memory is down", degrade)

fewer = {k: v for k, v in layers.items() if k not in ("memory service", "sandbox provider")}
report("fewer vendors (no separate memory, no sandbox)", fewer)

weakest = min(layers, key=layers.get)
print(f"\nWeakest link: {weakest} ({layers[weakest]:.2%} alone = {downtime(layers[weakest]):,.0f} min/month).")
print(f"Eight layers at 99.9% each would give {0.999 ** 8:.2%} end to end.")

Output:

Availability of one agent request path (all layers in series)

   configuration                                    avail.              downtime
   single provider, every layer required           98.806%         516 min/month
   + gateway failover to a 2nd model provider      99.300%         303 min/month
   + degrade gracefully if RAG or memory is down   99.498%         217 min/month
   fewer vendors (no separate memory, no sandbox)  99.003%         431 min/month

Weakest link: model provider (99.50% alone = 216 min/month).
Eight layers at 99.9% each would give 99.20% end to end.

Eight dependable-looking services add up to more than eight hours of monthly downtime in the request path. The biggest wins come from redundancy on the weakest link โ€” usually the model provider โ€” and from designing the agent to degrade gracefully rather than fail when an optional layer is down.

14 Market Structure

โ–ถ Two approaches Toggle to compare.

15 Consolidation, 2025โ€“2026

Money and acquisitions show where the market thinks value lies โ€” and which categories are becoming features of bigger platforms:

โ–ถ Deals and milestones Click an event.

16 Commoditizing vs Differentiating, Buyer Personas and Key Shifts

Becoming standardized / commoditized
Basic model API wrappers
Generic vector storage
Basic function calling / tool use
Baseline tracing and logging
Common RAG patterns
Standard sandbox environments
More strategic / differentiating
Domain-specific context and data
Proprietary workflows and agent logic
Permission and security architecture
Verification logic and task-specific evals
Durable state and long-term memory
Enterprise integrations, user experience and productization
๐Ÿง‘โ€๐Ÿ’ป
Application developer
  • Builds AI features into products
  • Needs speed, simplicity, good docs
  • Often uses managed services
๐Ÿข
Enterprise AI / platform team
  • Builds internal agent platforms
  • Needs security, governance, integration
  • Combines managed and self-hosted
๐Ÿš€
Startup building agents
  • Moves fast, iterates on product
  • Needs flexible, cost-effective stack
  • Open source plus managed services
๐Ÿ”ฌ
Research / prototyping team
  • Explores new ideas and capabilities
  • Needs flexibility and latest tools
  • Open source and experimental tools
โœ“From prompt libraries to stateful runtimes and agent frameworks with 1.0 stability guarantees.
โœ“Rapid growth of MCP and the official registry; MCP becomes stateless and enterprise-ready.
โœ“More long-running and durable workflows โ€” and a durable-execution funding boom.
โœ“Sandboxes and browser infrastructure become core agent infrastructure.
โœ“Observability and evaluation consolidate into data and infrastructure platforms.
โœ“AI security becomes part of mainstream security platforms; supply-chain risk becomes real.
โœ“Closer integration between model providers and harness infrastructure.

End-to-end: selected examples at each layer

1Agent appweb, mobile, enterprise
2Runtime / orchestratorOpenAI SDK, LangGraph, CrewAI, ADK
3Context / memoryMem0, Zep, Postgres, Redis
4RAG / retrievalLlamaIndex, Pinecone, Elasticsearch
5Tools / MCPMCP servers, GitHub, internal APIs
6Sandbox / browserE2B, Modal, Browserbase, Playwright
7Durable workflowTemporal, Trigger.dev, Inngest
8Observability / evalsLangSmith, Braintrust, Arize, Langfuse
9External systemsSaaS, databases, internal tools
10Model layerOpenAI, Anthropic, Google, open models
The harness market in one paragraph
By September 2026 the harness market has become a full infrastructure stack with its own leaders at each layer โ€” frameworks at 1.0 (Microsoft Agent Framework, LangGraph, ADK), MCP as the connectivity standard with a stateless 2026 specification and a registry of tens of thousands of servers, specialist vendors for retrieval, memory, browsers and sandboxes, and durable execution valued like core infrastructure (Temporal at $12.55B). It is also consolidating: observability tools are being absorbed into data platforms (Langfuse into ClickHouse, Helicone into Mintlify, W&B into CoreWeave) and AI-security startups into security giants. Commodity layers are getting cheaper and more interchangeable; the defensible value sits in domain context, workflows, permissions, verification and durable memory โ€” the parts only you can build. Choose portable standards (MCP, OpenTelemetry) and a small number of well-understood dependencies, and design for failure at every layer.

Next: Harness Case Studies โ€” OpenAI Codex, Claude Code auto mode and Gemini Deep Research taken apart.

โš ๏ธ Notes

The source infographic's categories and examples are sound; updates and refinements as of September 2026: (1) ownership changes it doesn't show โ€” Langfuse is now part of ClickHouse (January 2026), Helicone was acquired by Mintlify and put into maintenance mode (March 2026), Weights & Biases (Weave) is part of CoreWeave, and in security Lakera is part of Check Point and Protect AI part of Palo Alto Networks; (2) MCP Registry โ€” it is described as a central repository; it remains a preview, pre-GA service, and server counts vary widely with counting method; (3) LiteLLM โ€” listed without caveat; its March 2026 supply-chain compromise is a useful reminder that gateways concentrate credentials; (4) Microsoft Agent Framework is the 1.0 successor to AutoGen and Semantic Kernel, which the infographic doesn't mention. Availability figures in the script and widgets are illustrative inputs, not vendor SLAs; stack-builder scores are illustrative ratings.

๐Ÿ”— References