2026 年における AI エージェント最適化プラットフォームの選定ガイド

Most "AI agent" platforms watch your agent. Fewer of them change it. The gap between those two verbs is the whole category this post is about.

An agent ships, it runs, and some fraction of its answers are wrong in ways your eval suite never predicted. A dashboard tells you the score dropped. A good platform tells you why, grounded in the actual production traces. A platform that optimizes agents does the next thing: it tells you whether the fault is the agent or something else in your stack, proposes a specific change, and gives you a safe way to ship that change and watch the result. Observe, evaluate, and then improve, where improve means a deploy, not a Jira ticket.

2026 has been a consolidation year for this tooling. Langfuse was acquired by ClickHouse, OpenAI acquired Promptfoo and is shutting down its own hosted Evals product on November 30, Cisco moved to acquire Galileo, and Helicone was folded into Mintlify and put in maintenance mode. Braintrust, the best-funded holdout, raised $80 million of its own. So there are two honest questions now, not one: does the platform close the loop or just watch, and who is still independent enough that the loop, and your data, stay yours?

#What separates an agent-optimization platform from an observability tool

We wrote these criteria to favor closing the loop, because that is the category. If all you need is a dashboard, most of the tools below will do, and our own AI observability comparison is the better read. For actually improving an agent in production, these are what matter:

It closes the loop. The platform proposes a change grounded in production traces, and ships it through managed configuration with versioning, targeting, and rollback. A dashboard or an eval report is where most tools stop.

It sees the whole agent. An agent failure is often a retrieval, database, tool, or infrastructure failure that shows up as a bad answer. A platform that only stores LLM spans is optimizing against half the evidence, and cannot tell you when the prompt was never the problem.

Evaluation you do not ration. Online and offline evals without a per-score meter, so coverage is a quality decision rather than a billing one.

Open standards and independence. OpenTelemetry-native ingestion and portable data, so your instrumentation is not hostage to one vendor's roadmap, acquisition, or shutdown.

Safe rollout. Immutable versions, cohort targeting, weighted canaries, and rollback by moving a label, so shipping a change to a live agent is reversible.

One tell cuts through the marketing: watch which noun a platform optimizes. Some optimize your evals, the scorers and the judges, which is sharpening the ruler rather than the thing it measures. LangSmith's alignment tooling, for one, tunes the evaluator, not the app. Optimizing an agent means changing what runs in production, not tightening the grade it gets.

#The platforms, ranked

#1. Pydantic Logfire: best overall for agent optimization

Pydantic Logfire, from the team behind Pydantic Validation and Pydantic AI, is the platform that most completely closes the loop. It keeps the whole distributed trace in one OpenTelemetry-native product: browser, services, databases, model calls, tool calls, and infrastructure, with evaluation scores attached to the same timeline.

Its optimizer reads the runs that scored badly and the thousands around them, finds the pattern, and proposes one evidence-cited change. The part that matters most is upstream of the fix: because it sees the whole trace and not just the model spans, it tells you whether the problem is the agent at all, or a slow retrieval, a rate-limited upstream, or a database timeout showing up as a bad answer. You optimize the thing that is actually broken, instead of tuning a prompt against an infrastructure bug for a week. Then managed configuration ships the change: an agent's instructions, model, settings, and tool definitions live as versioned config with targeting and weighted rollout, and rollback is moving a label. That is what a self-improving agent actually takes: diagnose, propose, ship, roll back. The phrase is a slogan until one platform does all four over the whole trace.

Evaluation runs on the same traces you already emit, online and offline, with no separate per-score meter: a score is an observation, billed at the same flat rate as any span, with a generous free tier. Because it is OpenTelemetry-native, any framework that speaks OTel lights up the same surfaces, and Pydantic AI is wired in out of the box. Your coding agent can query the whole trace over the MCP server in PostgreSQL-compatible SQL and act on what it finds. And in a year when Langfuse went to ClickHouse and Promptfoo went to OpenAI, Logfire is the independent, open option: nothing here is getting absorbed or sunset.

Honest limitation: if all you want is evals and budget is no concern, Braintrust is the more extensible platform for teams on the bleeding edge of agent evals.

Best for: engineering teams optimizing production agents who want to know what to fix, ship the fix, and keep the full-stack context in one place, on open standards.

#2. Braintrust: best for eval-driven teams

Braintrust is the most established eval-first platform, and Loop is a real optimizer: describe a goal in natural language and it drafts scorers, builds eval datasets from your logs, and turns production failures into permanent eval cases. Customers cite real outcomes, including Notion's team going from three to thirty issues fixed a day. And it does ship the fix: a prompt registry fetched at runtime promotes a new version without a code deploy, gated on passing evals.

The limits are scope and openness. Braintrust optimizes the prompt and the LLM output; it does not manage the agent's whole configuration, and its own documentation says it focuses on LLM spans, so it cannot tell you when the real fault was the retrieval step or the database rather than the prompt. It runs on a proprietary datastore, self-hosting is Enterprise-only, and it puts a meter on evaluation itself: $1.50 to $2.50 per thousand scores depending on tier, charged on top of the model tokens each LLM-as-judge score already burns, plus processed data by the gigabyte. At production scale that per-score meter dominates. Three million scored runs a month is roughly $4,400 in score charges on the Pro plan, where the same three million on Logfire are just observations, a few dollars, with no score line at all. It closes a loop, a narrower one, on the AI output, beside the observability stack that holds the rest of your system.

Best for: teams whose bottleneck is eval coverage and CI/CD regression gates. Pricing: free tier; Pro at $249/month (Braintrust's docs say this drops to $100 in September 2026); Enterprise custom. Full breakdown: Logfire vs Braintrust.

#3. Arize AX: best if you are all-in on OpenInference

Arize markets AX as the platform for "self-improving agents," but a self-improving agent is one that closes the loop over the whole trace, which is exactly why Logfire sits above it here. Where Arize actually leads is elsewhere: OpenInference-native tracing through its Phoenix component, and a deep ML-monitoring lineage of drift detection and model-performance metrics. Its prompt-optimization is real: meta-prompting that generates an improved prompt from eval feedback, versions it in a Prompt Hub, and promotes it in a few clicks.

The tradeoffs are scope, lock-in, and cost. Arize comes from the ML-monitoring world and thinks in model metrics more than application traces; Phoenix is source-available under the Elastic License, not OSI open source; and the production optimization lives in the commercial AX product, whose dual-axis billing, charged per span and per gigabyte of payload, climbs fast on RAG workloads and runs well above Logfire's flat $2 per million. Phoenix also carries a recurring history of breaking upgrades that self-hosters absorb version to version, catalogued in Arize's own migration notes.

Best for: teams standardized on OpenInference and Phoenix, or ML teams that think in drift and model-performance metrics. Pricing: free tier; Pro from $50/month; Enterprise custom. Full breakdown: Logfire vs Arize.

#4. LangSmith: best for LangChain and LangGraph stacks

LangSmith offers strong tracing and evaluation with the tightest integration into LangChain and LangGraph, and a prompt hub that lets you version and update prompts without a full deploy. Its optimization is mostly a playground assistant and external cookbook recipes rather than an in-product optimizer, and its recent alignment tooling tunes the evaluator, not the app. It is framework-agnostic on paper but strongest inside that ecosystem, closed source, with self-hosting reserved for Enterprise and a per-seat plus per-trace price that, at scale, pushes teams to sample their traces down to a fraction just to control the bill, which is the opposite of what observability is for. Base retention is only 14 days; keeping traces the full 400 days costs about ten times as much per trace, and that ceiling applies even on your own hardware.

Best for: teams committed to LangChain and LangGraph. Pricing: free Developer tier; Plus from $39/seat/month; Enterprise custom. Full breakdown: Logfire vs LangSmith.

#5. Langfuse: best open-source option

Langfuse is a broad open-source platform, MIT-licensed, with tracing, prompt management, datasets, and evaluation, and it is the default choice for teams that have no budget for an AI observability tool and want to self-host and self-manage (ClickHouse acquired Langfuse in January 2026; it stays open source and self-hostable). It manages prompts and runs experiments, with textbook versioning and label-based rollout, but it observes and scores rather than proposing a trace-backed fix, so the optimizer half of the loop is one you assemble yourself. Self-hosting has its own recursive cost: you now run Postgres, ClickHouse, Redis, and object storage, infrastructure that itself needs watching, so you end up needing an observability solution to monitor your observability solution. And important governance features (project-level RBAC, SCIM, audit logs) require a commercial license key even when self-hosted.

Best for: devs who want to run infrastructure on their homelabs and teams that need a free and self-hostable platform and will build their own optimizer on top. Pricing: free self-host; cloud from $29/month. Full breakdown: Logfire vs Langfuse.

#6. The eval specialists: DeepEval, Promptfoo, Patronus

DeepEval, Promptfoo, and Patronus are excellent at the pre-ship half of the problem. DeepEval brings pytest-style LLM-as-judge testing and a real prompt optimizer, though one that tunes against your goldens rather than production traces. Promptfoo brings config-driven local evals and red-teaming. Patronus brings purpose-built judge models for hallucination detection and agent stress-testing. They are testing and evaluation frameworks, not production optimization platforms: they tell you what is wrong before or after a run, but they do not manage a live agent's configuration or ship a fix into production. Note the 2026 shakeout: Promptfoo is now owned by OpenAI, and OpenAI's own hosted Evals product is being shut down on November 30, though the open-source openai/evals repository is unaffected.

Best for: building a rigorous eval and red-teaming step into CI, upstream of whatever platform runs your loop.

#7. Elastic: a search company with an LLM dashboard

Elastic is a search and logging company that has bolted LLM observability onto its APM product. It will show you an agent's prompts, responses, token counts, and latency, and it markets AI agent optimization. Look at what backs that claim and it thins out fast. The optimization story is a short Elastic engineering blog series: engineers manually rewriting prompts to cut cost on internal workloads, not a product feature. The early-2026 Agent Builder builds retrieval agents over Elasticsearch data; it does not improve them. And the LLM instrumentation is a recent integration layer on a search engine, not tracing designed for the shape of an agent run. Elastic is also heavy to operate in the first place: between the query DSL, index lifecycle management, and cluster tuning, running it well can feel like it takes a PhD in Elastic, a reputation Elastic half-conceded by shipping AutoOps to catch misconfigurations.

So Elastic shows you prompts. It does not manage them. There is no versioning, no targeting, no rollout, no rollback, no optimizer that proposes a fix, and nothing that turns a bad run into a shipped change. It is an observability view of an agent sold next to the words "agent optimization," and it is last here because a blog post is not a product.

Best for: teams already deep in Elastic who want LLM calls on the same dashboards as the rest of their logs, and who will do any actual optimizing somewhere else.

#Comparison at a glance

Platform

Closes the loop (propose + ship)?

Trace scope

Open standards

Eval / score pricing

Best for

Pydantic Logfire

Yes: diagnoses fault, optimizer proposes, managed config ships

Whole distributed trace

OTel-native, independent

No score meter, flat spans

Optimizing agents in full production context

Braintrust

Yes, for prompts (Loop + registry)

LLM spans (OTel ingest)

Proprietary datastore (Brainstore)

Per-score ($1.50-2.50/1k) + per-GB

Eval-driven CI/CD teams

Arize AX

Yes: meta-prompt optimizer + Prompt Hub

LLM / agent spans

Phoenix ELv2; AX proprietary

Dual-axis: spans + $/GB payload

Enterprise ML + LLM teams

LangSmith

Partial: prompt hub + assistant

Agent / LLM traces

Closed

Seats + traces

LangChain / LangGraph shops

Langfuse

No built-in optimizer

LLM traces (OTel-compatible)

Open source (MIT)

Units include scores

Self-hosted, OSS-first teams

Eval specialists

No: test and score pre-ship

Test-time, not production

Mostly OSS

Varies

Rigorous CI eval and red-teaming

Elastic

No: shows prompts, manages nothing

LLM dashboards bolted on search

OTel; LLM layer bolted on

n/a

Existing Elastic customers

#How to choose

Start with your bottleneck. If you cannot see why the agent failed, or cannot tell whether the agent or your infrastructure is at fault, you need whole-trace observability first, and any full-stack option beats an LLM-only one. If you can see the failures but cannot turn them into shipped fixes without a deploy cycle, you need a platform that closes the loop, and that is where Logfire, Braintrust's Loop, and Arize AX actually compete. If you need to own your infrastructure, start open-source with Langfuse or self-hosted Phoenix. If your problem is pre-ship regression and red-teaming, add an eval specialist to CI regardless of which platform runs your production loop.

For most teams improving a real agent in production on an open, polyglot stack, Pydantic Logfire is the strongest starting point: it is the one that tells you what to fix, proposes the fix, and ships it, over the whole trace, without a per-score meter, and without asking you to bet on a single vendor's proprietary format in a year when everyone else is getting acquired.

#Frequently asked questions

What is AI agent optimization?

It is the loop that turns a failing production run into a shipped improvement: capture the trace, evaluate it, work out whether the agent or the surrounding system is at fault, propose a specific change to the agent's prompt or configuration, ship that change with a safe rollout, and measure the result. Optimization is distinct from observability, which stops at showing you what happened.

What is the best Braintrust alternative for agent optimization?

Pydantic Logfire. Braintrust's Loop is strong, and it does ship prompt changes, but it optimizes the LLM output inside Braintrust. Logfire works over the whole distributed trace, so it can tell you when the real problem was the retrieval or the database rather than the prompt, ships the agent's whole configuration through managed rollout, and does it on OpenTelemetry with no per-score meter.

Does Elastic do AI agent optimization?

Not really. Elastic shows you an agent's prompts and traces and markets AI agent optimization, but the optimization is a handful of engineering blog posts, not a product feature: engineers manually rewriting prompts to cut cost. Elastic displays prompts; it does not version, target, roll out, or roll back anything, and it has no optimizer that proposes a fix. It is an observability view of an agent, not a platform that improves one.

Are there open-source agent-optimization options?

Langfuse is MIT-licensed and self-hostable, and Arize Phoenix is source-available under the Elastic License. Both give you observability and evaluation to build a loop on, rather than a managed optimizer. Because Pydantic Logfire is OpenTelemetry-native, your instrumentation stays portable regardless of where you run it.

What happened to the AI eval tooling in 2026?

It consolidated fast. ClickHouse acquired Langfuse in January, OpenAI acquired Promptfoo in March and announced its own hosted Evals product will shut down on November 30, and Cisco moved to acquire Galileo. The open-source openai/evals framework on GitHub is unaffected, but the pattern is the point: build your loop on a platform that is OpenTelemetry-native and independent, so the loop and your data stay yours no matter who gets bought next.

#Try Pydantic Logfire free

You can have whole-stack agent traces, online and offline evaluation, and the optimization loop running in a few minutes. The free tier includes ten million spans a month.

Start free with Pydantic Logfire

AI is still just engineering.