LangGraph+25 Positive

Operator Insights

What are teams learning from building on LangGraph and running it in production?

14 accounts: 2 left it or turned it down, 1 evaluated it and 11 adopted it

Left it or turned it down2

Nicola RichliBuilder of yeos.aiyeos.ai
yeos.ai's enterprise AI platform, serving several customers whose requests were mostly sequential.
DEV Community, published Jul 27, 2026Read the write-up (opens DEV Community in a new tab)
Observation
yeos.ai dropped LangGraph from its enterprise AI platform after months in use, once the team spent more time on the framework than on its own business logic.
Takeaway
Before you adopt a graph, count how many requests actually branch or wait. If most run in a straight line, plain components are cheaper to debug.
What happened
yeos.ai built its platform's agent orchestration on LangGraph and used it for months (DEV Community post, Jul 27, 2026). As customers grew, a bug spanning many state transitions became harder to trace than a bug in a function, and most requests never branched. The team dropped LangGraph completely for small composable components that are easier to test.
“Over time we found ourselves thinking more about LangGraph than about our own business logic.”
Nicola Richli
- Public practitioner write-up (DEV Community)
- No team metrics reported
Deployment details
yeos.ai's enterprise AI platform, serving several customers whose requests were mostly sequential.
- Public practitioner write-up (DEV Community)
- No team metrics reported
Team
yeos.ai engineering
Duration
Months, ending before Jul 27, 2026
Workload
Agent orchestration for an enterprise AI platform

James Briggs (jamesbriggs)Oct 8, 2025Hacker NewsAurelio Labs left LangChain, LangGraph and the OpenAI Agents SDK for its own framework after using all three extensively.
James Briggs (jamesbriggs)Maintainer of the graphai frameworkAurelio AI
Aurelio Labs builds agents for its own products and maintains graphai, an open-source async agent framework.
Hacker News, published Oct 8, 2025Read the write-up (opens Hacker News in a new tab)
Observation
Aurelio Labs left LangChain, LangGraph and the OpenAI Agents SDK for its own framework after using all three extensively.
Takeaway
A team that knows the exact shape of its workflows can outgrow a general framework. Keep model calls and tools behind your own interfaces so the move stays cheap.
What happened
Briggs wrote on Hacker News on Oct 8, 2025 that his team had used LangChain, LangGraph and OpenAI's Agents SDK pretty extensively, and now rolls its own, which he calls a much better and cleaner experience.
Public practitioner write-up (Hacker News comment)
No team metrics reported
Deployment details
Aurelio Labs builds agents for its own products and maintains graphai, an open-source async agent framework.
Public practitioner write-up (Hacker News comment)
No team metrics reported
Team
Aurelio Labs engineering
Workload
Agent orchestration for Aurelio Labs' products

Evaluated it1

Samuel ColvinFounderPydantic
Pydantic needed AI features for Logfire, its observability product, in late 2024. Colvin now leads Pydantic AI, which competes with LangGraph.
Software Engineering Daily, published Dec 4, 2025Listen to the episode (opens Software Engineering Daily in a new tab)
Observation
Pydantic evaluated LangGraph, CrewAI and other frameworks for Logfire's AI features, found them not type safe, and built Pydantic AI instead.
Takeaway
If type safety matters to your team, test how a framework types its state and tool calls before you commit. Weigh this account knowing its speaker sells an alternative.
What happened
Colvin expected one of the frameworks built on Pydantic, LangChain, LangGraph, CrewAI and LlamaIndex among them, to fit Logfire (Software Engineering Daily, Dec 4, 2025). He evaluated them, found them not type safe and short of the engineering standard Pydantic holds its own libraries to, and the team built Pydantic AI.
“Started looking at them and was super disappointed by what I found. They're not type safe. I think type safety is incredibly important and only getting more important with AI's writing code.”
Samuel Colvin, at 13:08 in the recording
- Podcast episode
- No team metrics reported
Deployment details
Pydantic needed AI features for Logfire, its observability product, in late 2024. Colvin now leads Pydantic AI, which competes with LangGraph.
- Podcast episode
- No team metrics reported
Team
Pydantic Logfire and Pydantic AI engineering
Workload
AI features for Logfire
Product version
LangGraph and CrewAI as of late 2024

Adopted it11

Mikolaj WawrzyniakAuthor of the GitLab engineering articleGitLab
GitLab Duo Agent Platform, whose agent flows for GitLab engineers and customers compile from YAML into LangGraph graphs.
GitLab blog, published Sep 22, 2026Read the write-up (opens GitLab blog in a new tab)
Observation
GitLab chose LangGraph as the foundation of GitLab Duo Agent Platform, then built a YAML layer on top that cut its own code per agent flow by 45%.
Takeaway
Use LangGraph as the runtime and give flow authors a higher-level definition on top, so each new flow is a configuration file that compiles into the graph.
What happened
GitLab researched agent frameworks and selected LangGraph for durable execution and traceability (GitLab blog, Sep 22, 2026). Hand-wired graphs ran to 450 or more lines of Python per flow, which GitLab found too low-level to scale. Its Flow Registry now compiles YAML flow definitions into LangGraph flows, and GitLab says that reduced its own code per agentic flow by 45%.
“We researched available AI frameworks and selected LangGraph, an agent runtime and low-level orchestration framework from LangChain, as the foundation for GitLab Duo Agent Platform.”
GitLab
- Public engineering write-up
- Team metrics reported
Deployment details
GitLab Duo Agent Platform, whose agent flows for GitLab engineers and customers compile from YAML into LangGraph graphs.
- Public engineering write-up
- Team metrics reported
Team
GitLab Duo Agent Platform team
Workload
Agent flows for GitLab engineers and Duo Agent Platform customers
Product version
LangGraph under GitLab's Flow Registry
Measurement
45% less of GitLab's own code per agentic flow, self-reported

WayfairJan 15, 2026Wayfair Tech BlogWayfair runs Wilma, its supplier-transfer ticket flagger, as an event-driven LangGraph workflow that beat its human targets with 95% precision and 97% recall.
Farshid BahramiCo-author, Wayfair Tech BlogWayfair
Wayfair customer service: every new comment on a supplier-transfer ticket triggers the workflow, which decides in about one second whether an associate must step in.
Wayfair Tech Blog, published Jan 15, 2026Read the write-up (opens Wayfair Tech Blog in a new tab)
Observation
Wayfair runs Wilma, its supplier-transfer ticket flagger, as an event-driven LangGraph workflow that beat its human targets with 95% precision and 97% recall.
Takeaway
Decide what code can decide before the model runs, and give each kind of input its own prompt.
What happened
Wilma runs on an event-driven pipeline using LangGraph (Wayfair Tech Blog, Jan 15, 2026). The graph identifies the comment author in code before any model call, then classifies with an author-specific prompt. Against target baselines of 80% precision and 90% recall, Wilma reached 95% precision and 97% recall, cut time to resolution 29% against the period before it, and cleared about 65% of tickets without a person in early results. Wayfair monitors it in Arize.
“Wilma significantly outperformed its original target human baselines of 80% precision and 90% recall, achieving 95% precision, 97% recall, 99.8% reason accuracy and 99.7% timely flagging…”
Wayfair Tech Blog
Public engineering write-up
Team metrics reported
Deployment details
Wayfair customer service: every new comment on a supplier-transfer ticket triggers the workflow, which decides in about one second whether an associate must step in.
Public engineering write-up
Team metrics reported
Team
Wayfair machine learning and customer service technology
Workload
Flagging supplier-transfer ticket comments that need an associate
Product version
Event-driven LangGraph workflow, monitored in Arize
Measurement
95% precision and 97% recall against 80% and 90% targets; 29% faster resolution
GrabMar 19, 2026Grab Tech BlogGrab's Analytics Data Warehouse team, which supports more than 1,000 users a month, runs a LangGraph multi-agent system behind FastAPI to answer their data questions.
Sneh AgrawalCo-author, Grab Tech BlogGrab
Grab's Analytics Data Warehouse team, which supports over 1,000 users a month across more than 15,000 tables.
Grab Tech Blog, published Mar 19, 2026Read the write-up (opens Grab Tech Blog in a new tab)
Observation
Grab's Analytics Data Warehouse team, which supports more than 1,000 users a month, runs a LangGraph multi-agent system behind FastAPI to answer their data questions.
Takeaway
Point agents at the repeated questions that eat a platform team's week, and keep the hard ones collaborative.
What happened
The team uses FastAPI to handle requests and LangGraph to manage the state and loops of multi-agent collaboration (Grab Tech Blog, Mar 19, 2026). Agents answer simpler questions on their own and work together on complex ones. Grab says the system freed hundreds of hours of engineering time a month, and first answers reach Slack within minutes where manual searches took hours.
“FastAPI & LangGraph: We use FastAPI to handle requests and LangGraph to manage the complex state and cyclical logic required for multi-agent collaboration.”
Grab Tech Blog
Public engineering write-up
Team metrics reported
Deployment details
Grab's Analytics Data Warehouse team, which supports over 1,000 users a month across more than 15,000 tables.
Public engineering write-up
Team metrics reported
Team
Grab Analytics Data Warehouse team
Workload
Data questions to a team that supports more than 1,000 users a month
Product version
LangGraph behind FastAPI, with PostgreSQL memory
Measurement
Hundreds of engineering hours freed a month, self-reported
TiendanubeApr 28, 2026Tadeo Donegana Braunschweig's blogTiendanube runs its merchant copilot on LangGraph at more than 120,000 conversations a week, and had to build its own cleanup because the open-source Postgres checkpointer never expires a checkpoint.
Tadeo Donegana BraunschweigAI engineer on the copilot teamTiendanube
Tiendanube's conversational copilot, which lets merchants manage their stores in natural language. A LangGraph supervisor routes to stateless specialist subgraphs.
Tadeo Donegana Braunschweig's blog, published Apr 28, 2026Read the write-up (opens Tadeo Donegana Braunschweig's blog in a new tab)
Observation
Tiendanube runs its merchant copilot on LangGraph at more than 120,000 conversations a week, and had to build its own cleanup because the open-source Postgres checkpointer never expires a checkpoint.
Takeaway
Plan checkpoint retention before launch: partition the tables, delete old threads on a schedule, and checkpoint only the graph that needs to resume.
What happened
An average conversation, two messages and three tool calls, leaves about 93 records across four checkpoint tables, and one week of staging tests grew checkpoint_blobs to 56 MB (author's blog, Apr 28, 2026). The open-source Postgres checkpointer has no expiry, so the team creates the tables through its own Alembic migrations, deletes old threads with a cron job over pg_partman partitions, exports messages weekly to S3, and checkpoints only the supervisor graph.
Public engineering write-up (author's blog)
Team metrics reported
Deployment details
Tiendanube's conversational copilot, which lets merchants manage their stores in natural language. A LangGraph supervisor routes to stateless specialist subgraphs.
Public engineering write-up (author's blog)
Team metrics reported
Team
Tiendanube merchant copilot team
Workload
More than 120,000 merchant conversations a week
Product version
Open-source LangGraph with the Postgres checkpointer
Measurement
About 93 checkpoint records per conversation; 56 MB of checkpoint_blobs after a week of staging
AirbnbSep 29, 2026Interrupt NYC 2026 (LangChain's YouTube channel)Airbnb's trust agents on LangGraph can answer "I don't know", pause for days with interrupts, and keep case data outside the graph.
Pedro RodriguezDirector of Engineering, MLAirbnb
Trust and safety decisions on guest and host cases. Cases that wait days or weeks for signals or a person's decision. A confident wrong answer is the costliest outcome.
Interrupt NYC 2026 (LangChain's YouTube channel), published Sep 29, 2026Watch the talk (opens Interrupt NYC 2026 (LangChain's YouTube channel) in a new tab)
Observation
Airbnb's trust agents on LangGraph can answer "I don't know", pause for days with interrupts, and keep case data outside the graph.
Takeaway
Give every agent a way to say it does not know, and decide obvious cases in plain code before the model runs. Keep long-lived case data in your own database, outside graph state.
What happened
Airbnb trust agents return structured output with an abstain option that sends the case to a person. The first version abstained on about 25% of cases it should approve. Splitting policy selection from policy application reduced abstention and raised cost. Deterministic rules approve, deny, or escalate obvious cases before any LLM call, while small subagents feed a synthesizer. interrupt() pauses cases for days until a signal or decision arrives. Case data stays in a separate database. Rodriguez reports that Airbnb chose LangGraph after trying other frameworks, for data access and access controls, stop and replay, and per-run inspection. It also spared engineers from operating a custom framework.
Vendor customer story (talk at LangChain's Interrupt NYC 2026, on LangChain's YouTube channel)
Team metric reported
Deployment details
Trust and safety decisions on guest and host cases. Cases that wait days or weeks for signals or a person's decision. A confident wrong answer is the costliest outcome.
Vendor customer story (talk at LangChain's Interrupt NYC 2026, on LangChain's YouTube channel)
Team metric reported
Team
Airbnb's trust and safety AI team
Workload
Policy decisions on guest and host trust cases, some waiting days or weeks
Product version
LangGraph with interrupts, structured outputs, and MCP for data access
Measurement
Abstain rate of about 25% in the first version, lower after splitting the work
ToyotaJul 15, 2026Interrupt 26 (LangChain's YouTube channel)Toyota runs more than 50 ToyotaGPT agents in production on dynamically generated LangGraph graphs, each agent a config file.
Ravi Chandu Ummadisetti and Kordel FranceToyota enterprise AI teamToyota
ToyotaGPT, Toyota's enterprise agent platform.
Interrupt 26 (LangChain's YouTube channel), published Jul 15, 2026Watch the talk (opens Interrupt 26 (LangChain's YouTube channel) in a new tab)
Observation
Toyota runs more than 50 ToyotaGPT agents in production on dynamically generated LangGraph graphs, each agent a config file.
Takeaway
Once several agents share a shape, generate their graphs from configuration so a new agent only needs its settings reviewed.
What happened
In a talk at LangChain's Interrupt 26 conference (video published Jul 15, 2026), Toyota's enterprise AI team described ToyotaGPT's dynamic LangGraph graph generation, with more than 50 agents in production, every one a config file. LangChain's description of the talk says the platform cut agent delivery from six months and six engineers to four days and one engineer.
Vendor customer story (talk at LangChain's Interrupt 26, on LangChain's YouTube channel)
Team metrics reported
Deployment details
ToyotaGPT, Toyota's enterprise agent platform.
Vendor customer story (talk at LangChain's Interrupt 26, on LangChain's YouTube channel)
Team metrics reported
Team
Toyota's enterprise AI team
Workload
50+ production agents on ToyotaGPT
Product version
Dynamically generated LangGraph graphs, one config file per agent
Measurement
Agent delivery from six months and six engineers to four days and one engineer
LATAM AirlinesJun 30, 2026Interrupt 26 (LangChain's YouTube channel)LATAM Airlines runs its Concierge travel agent on LangGraph, with a supervisor over a tool-per-agent design, and restructured it for 15% lower cost.
Nico Venegas and Claudio Urbina LaraLATAM Airlines AI teamLATAM Airlines
LATAM's Concierge agent for travelers, with about 4,000 daily users.
Interrupt 26 (LangChain's YouTube channel), published Jun 30, 2026Watch the talk (opens Interrupt 26 (LangChain's YouTube channel) in a new tab)
Observation
LATAM Airlines runs its Concierge travel agent on LangGraph, with a supervisor over a tool-per-agent design, and restructured it for 15% lower cost.
Takeaway
Put a supervisor in front of narrow agents, trace every route, and revisit the structure once real traffic shows where the cost goes.
What happened
In a talk at LangChain's Interrupt 26 conference (video published Jun 30, 2026), LATAM described Concierge, a LangGraph-based travel agent with a tool-per-agent architecture under a supervisor, which LangSmith let it restructure for a 15% cost cut. A 13% out-of-scope rate turned out to be a product gap.
Vendor customer story (talk at LangChain's Interrupt 26, on LangChain's YouTube channel)
Team metrics reported
Deployment details
LATAM's Concierge agent for travelers, with about 4,000 daily users.
Vendor customer story (talk at LangChain's Interrupt 26, on LangChain's YouTube channel)
Team metrics reported
Team
LATAM Airlines AI team
Workload
Concierge travel agent, about 4,000 daily users
Product version
LangGraph supervisor over one agent per tool, traced in LangSmith
Measurement
15% lower cost after restructuring
QodoMar 21, 2025Qodo blogQodo took LangGraph from a first prototype to its production coding agent, and now tunes how structured each flow is as models improve.
Sagi MedinaEngineer and author of the postQodo
Qodo's coding agent, built from graph steps it reuses across flows. About four months from framework choice to the write-up.
Qodo blog, published Mar 21, 2025Read the write-up (opens Qodo blog in a new tab)
Observation
Qodo took LangGraph from a first prototype to its production coding agent, and now tunes how structured each flow is as models improve.
Takeaway
Treat graph density as a setting you will change: sparse, structured graphs for today's models, more open graphs as models improve, with shared nodes so new flows cost little.
What happened
Qodo wanted to move from rigid, predefined coding flows to an agent that adapts to any request while keeping its opinionated workflow. The team chose LangGraph for a first prototype and found it mature enough to carry the agent to production. Context collection and validation nodes are reused across flows such as TDD, and persistence to Postgres took a few lines with a checkpointer. The weak spots were documentation that lagged fast releases and the difficulty of testing flows that depend on the IDE.
“…we settled on LangGraph for our initial proof of concept. We were pleasantly surprised to see that the framework has proven flexible and mature enough to carry us all the way to production.”
Sagi Medina
Public engineering write-up
No team metrics reported
Deployment details
Qodo's coding agent, built from graph steps it reuses across flows. About four months from framework choice to the write-up.
Public engineering write-up
No team metrics reported
Team
Qodo engineering team building its coding agent
Duration
About four months from choosing LangGraph to the March 2025 write-up
Workload
Coding agent flows: context collection, planning, code generation, and validation with tests
Product version
LangGraph with the PostgresSaver checkpointer
monday.comFeb 18, 2026LangChain blogmonday.com built its AI service workforce as a LangGraph-based ReAct agent and runs every CI build as a LangSmith experiment before changes reach production.
Gal Ben AriehGroup Tech Leadmonday.com
monday service's AI service workforce, an agent that resolves inquiries across enterprise service management use cases.
LangChain blog, published Feb 18, 2026Read the write-up (opens LangChain blog in a new tab)
Observation
monday.com built its AI service workforce as a LangGraph-based ReAct agent and runs every CI build as a LangSmith experiment before changes reach production.
Takeaway
Run agent evals in CI on every change, one experiment per run, so a regression shows which step diverged before release.
What happened
LangChain's customer story of Feb 18, 2026 says the agent is a customizable, LangGraph-based ReAct agent. monday.com logs each CI run as a distinct LangSmith experiment, with each test suite as a dataset, and parallelized Vitest files and LLM evaluations to speed the suite up.
“AI service workforce is a customizable, LangGraph-based, ReAct agent, designed to automate and resolve inquiries across any enterprise service management use case.”
Gal Ben Arieh
Vendor customer story (LangChain blog)
No team metrics reported
Deployment details
monday service's AI service workforce, an agent that resolves inquiries across enterprise service management use cases.
Vendor customer story (LangChain blog)
No team metrics reported
Team
monday service engineering
Workload
Service-management inquiries resolved by an agent
Product version
LangGraph ReAct agent, evaluated in LangSmith from Vitest
LinkedInSep 10, 2025LinkedIn Engineering BlogLinkedIn built its agent platform, the one behind Hiring Assistant, on LangGraph adapted to its own messaging and memory infrastructure.
Karthik RamgopalDistinguished Engineer, generative AI applications and platformLinkedIn
LinkedIn's generative AI platform and Hiring Assistant, a supervisor and sub-agent recruiting agent. LinkedIn's messaging system carries agent traffic.
LinkedIn Engineering Blog, published Sep 10, 2025Read the write-up (opens LinkedIn Engineering Blog in a new tab)
Observation
LinkedIn built its agent platform, the one behind Hiring Assistant, on LangGraph adapted to its own messaging and memory infrastructure.
Takeaway
Adopt the framework behind your own platform layer, and handle workflows that need no judgment in procedural code.
What happened
LinkedIn's engineering blog says it embraced LangGraph for its integrations, production readiness and observability, and adapted it to LinkedIn messaging and memory with custom-built providers (Sep 10, 2025). Agents register in a central skill registry, LangSmith traces them before production and OpenTelemetry in production. In a QCon London 2025 talk, Ramgopal and Daniel Hewlett stated that LinkedIn picked LangChain and LangGraph behind its own adapters so it can switch frameworks relatively cheaply.
Public engineering write-up (LinkedIn Engineering Blog, with the QCon London 2025 talk on InfoQ)
No team metrics reported
Models: Azure OpenAI and fine-tuned open models
Deployment details
LinkedIn's generative AI platform and Hiring Assistant, a supervisor and sub-agent recruiting agent. LinkedIn's messaging system carries agent traffic.
Public engineering write-up (LinkedIn Engineering Blog, with the QCon London 2025 talk on InfoQ)
No team metrics reported
Models: Azure OpenAI and fine-tuned open models
Team
LinkedIn's generative AI platform team and the Hiring Assistant team
Workload
Hiring Assistant recruiting agent and LinkedIn's shared agent platform
Product version
LangGraph with LinkedIn's messaging and memory providers
Model
Azure OpenAI models and fine-tuned open-source models hosted on-premises, per the QCon talk
BlackRockJun 12, 2025Interrupt 2025 (LangChain's YouTube channel)BlackRock built the agents behind its Aladdin platform on LangGraph, with a plugin registry through which more than 50 engineering teams contribute tools and agents.
Brennan Rosales and Pedro Vicente ValdezBlackRock AI engineering teamBlackRock
BlackRock's Aladdin platform, which serves US$11 trillion in assets across more than 100 applications.
Interrupt 2025 (LangChain's YouTube channel), published Jun 12, 2025Watch the talk (opens Interrupt 2025 (LangChain's YouTube channel) in a new tab)
Observation
BlackRock built the agents behind its Aladdin platform on LangGraph, with a plugin registry through which more than 50 engineering teams contribute tools and agents.
Takeaway
When many teams contribute tools to one agent platform, register them in one place and run the evaluations in CI every day.
What happened
In a talk at LangChain's Interrupt 2025 conference (video published Jun 12, 2025), BlackRock's AI engineering team described the agentic architecture behind Aladdin, built on LangGraph, with a plugin registry that lets more than 50 engineering teams contribute tools and agents, and evaluation-driven development tested daily in CI/CD.
Vendor customer story (talk at LangChain's Interrupt 2025, on LangChain's YouTube channel)
No team metrics reported
Deployment details
BlackRock's Aladdin platform, which serves US$11 trillion in assets across more than 100 applications.
Vendor customer story (talk at LangChain's Interrupt 2025, on LangChain's YouTube channel)
No team metrics reported
Team
BlackRock AI engineering, with tools from more than 50 engineering teams
Workload
Agents in the Aladdin platform
Product version
LangGraph, with a plugin registry for tools and agents

Left it or turned it down2

yeos.ai dropped LangGraph from its enterprise AI platform after months in use, once the team spent more time on the framework than on its own business logic.

Aurelio Labs left LangChain, LangGraph and the OpenAI Agents SDK for its own framework after using all three extensively.

Evaluated it1

Pydantic evaluated LangGraph, CrewAI and other frameworks for Logfire's AI features, found them not type safe, and built Pydantic AI instead.

Adopted it11

GitLab chose LangGraph as the foundation of GitLab Duo Agent Platform, then built a YAML layer on top that cut its own code per agent flow by 45%.

Wayfair runs Wilma, its supplier-transfer ticket flagger, as an event-driven LangGraph workflow that beat its human targets with 95% precision and 97% recall.

Grab's Analytics Data Warehouse team, which supports more than 1,000 users a month, runs a LangGraph multi-agent system behind FastAPI to answer their data questions.

Tiendanube runs its merchant copilot on LangGraph at more than 120,000 conversations a week, and had to build its own cleanup because the open-source Postgres checkpointer never expires a checkpoint.

Airbnb's trust agents on LangGraph can answer "I don't know", pause for days with interrupts, and keep case data outside the graph.

Toyota runs more than 50 ToyotaGPT agents in production on dynamically generated LangGraph graphs, each agent a config file.

LATAM Airlines runs its Concierge travel agent on LangGraph, with a supervisor over a tool-per-agent design, and restructured it for 15% lower cost.

Qodo took LangGraph from a first prototype to its production coding agent, and now tunes how structured each flow is as models improve.

monday.com built its AI service workforce as a LangGraph-based ReAct agent and runs every CI build as a LangSmith experiment before changes reach production.

LinkedIn built its agent platform, the one behind Hiring Assistant, on LangGraph adapted to its own messaging and memory infrastructure.

BlackRock built the agents behind its Aladdin platform on LangGraph, with a plugin registry through which more than 50 engineering teams contribute tools and agents.