Managed agents
forany harness

Agent infrastructure to enable the right harness, model, and context for every task.

App:

1import { Checkfu } from "@checkfu/sdk"
2
3const checkfu = new Checkfu({
4  apiKey: process.env.CHECKFU_API_KEY!,
5})
6
7// 01 Create specialized architect and coding agents on different harnesses
8const architect = await checkfu.agents.create({
9  name: "lead-architect",
10  harness: "claude-code",
11  model: "claude-opus-5",
12  system: "Plan refactors, enforce architectural invariants, and review PR diffs.",
13  tools: [{ type: "agent_toolset_20260401" }],
14})
15
16const coder = await checkfu.agents.create({
17  name: "code-executor",
18  harness: "codex",
19  model: "gpt-5",
20  system: "Execute code modifications inside the sandbox and verify unit tests.",
21  tools: [{ type: "agent_toolset_20260401" }],
22})

Harness-model fit

Choosing the right model is only part of the equation. Pairing it with the right harness, context, and taste is what makes an agent actually work.

Claude Agent SDK

Codex app server

OpenCode

Cursor

OpenClaw

Hermes

Bring your own

Compute that fits the job, sessions that outlive it

Some agents need a real computer with a browser that stays up for days. Others need a fresh sandbox for one second of code. Pick per agent. The agent loop runs apart from the machine, so a lost worker never loses the session.

Durable execution

A worker can die mid-command. The run keeps going, and the result lands once.

A real computer

A desktop, a browser, and files that stay put for days.

A fast sandbox

Fresh compute in under a second. Run one job, then throw it away.

Mount files, repos, memory, and skills

Give your agents the exact context they need to act accurately. Stage Git repositories, stream files into the filesystem, retain cross-session memory, and load modular skills.

GitHub

acme/api

Files

q3-board-deck.pdf

Memory

maya-notes

Skills

meeting-notes

Fizz

agent 1

claude-code

Honey

agent 2

codex

Bumble

agent 3

opencode

Each agent mounts only what its job needs.

100+ native connectors & MCPs

Equip your agents with hundreds of tools across communication, code, research, data, and business automation — with vaulted OAuth, credential encryption, and human approval gates.

Work & Communication

Team chat, workspaces, calendar schedules, and email automation.

Code & Product

Issue tracking, design tokens, error monitoring, and code repositories.

Web & Research

Deep research, web crawling, neural search, and live browsers.

Data & Infrastructure

Cloud databases, serverless platforms, edge compute, and metrics.

Business & Automation

Payments, CRM pipelines, storefront orders, and workflow triggers.

And hundreds more through MCP & REST

Checkfu natively speaks the Model Context Protocol. Connect any custom MCP server, OpenAPI endpoint, or private microservice with zero protocol boilerplate.

Durable sessions on any channel : Slack, Discord, Telegram, WhatsApp, Microsoft Teams, or your own over the API

Deploy agents directly into team chat channels or connect via API. Every turn, tool call, and state transition is written to an append-only event ledger that survives crashes and exports to JSON.

Sessions/sesn_01J9A3RXK7QF2M8VZ0T4HDN6CE

sesn_01J9A3RXK7QF2M8VZ0T4HDN6CE

Idle

Fizzclaude-codesupport-desk-env#supportthread51s38.4k / 2.1k$0.192 hours ago

1.00×

maya09:22:31via slack #support

Acme’s nightly export has been sitting in `queued` since 02:10 and their CSM is asking for an ETA. Can you find out what’s wrong and get it moving?

Fizz09:22:31claude-code

Thought for 2.1s

Starting with the queue and the error tracker before I touch anything.

Thought for 3.4s

runner-11 claimed all three jobs at 02:10 and stopped heartbeating at 02:14, but its claim was never released, so the queue still counts the jobs as in flight. Confirming from the runner itself.

Thought for 4.0s

Root cause: a stale claim from a dead runner, not a bug in the export itself. Releasing the claim and requeuing the three jobs is a write against production, so it needs a human.

Thought for 1.8s

Found it, and it’s moving.

What happened. runner-11 claimed Acme’s three export jobs at 02:10, stopped heartbeating at 02:14 (Sentry EXPORT-412), and its claim was never released, so the queue kept treating the jobs as in flight. The export itself is fine.

What I did. With maya’s approval I released the claims and requeued the three jobs. One has completed, one is running, and the last should land in about ten minutes. I posted the same update in the #support thread for the CSM.

I wrote the diagnosis to the support runbook under exports/stuck-queued, so next time this is a one-query check.

Reply in the thread if the CSM needs a written incident note.

Idle · replied in #support

Session

IDsesn_01J9A3RXK7QF2M8VZ0T4HDN6CE

StatusIdle

Workspacewrkspc_acme-prod

Channel

AgentFizz agent_fizz

Harnessclaude-code on frontier-lg

Environmentsupport-desk-env

Vaultsacme-replica-ro, sentry, slack

Memorysupport-runbook

CreatedThu, Jul 16 · 09:22:28

UpdatedThu, Jul 16 · 09:23:19

Run

Runrun_7f31

Attempts

Approvals

Ledger38 events · exportable as JSON

Cost

Model (38.4k / 2.1k)$0.17Sandbox compute$0.02Total$0.19

Build AI-native products

From coding orchestrators and personal assistants to team coworkers and autonomous agents, build what’s next on Checkfu.

acme/api4 workspaces

fix-flaky-tests

cloud+9 −3

bump-deps

local+41 −41

migrate-api

cloud+212 −87

rename-config

local

local runner

fix-flaky-testsPR ready

CI is flaky on the payments suite. Find the leak and fix it.

running vitest --retry 2

3 failed, isolating test order

patch applied to test/setup.ts

412 passed, opening summary

A shared fixture leaked between files. Reset it per test; suite is green.

changes2

test/setup.ts+8 −2

vitest.config.ts+1 −1

Coding Orchestrator

Fleets of parallel coding agents running inside isolated cloud microVMs with git worktrees and automated pull request generation.

Multi-agent dispatch with architect and coder roles

Ephemeral git worktrees & sparse repository checkouts

Automated test runs and feedback repair loops

Human approval gates before applying code changes

Personal Assistant

A 24/7 personal executive assistant with long-term memory, daily schedule triage, and proactive morning briefs.

Persistent scoped memory store across sessions

Google Calendar & Gmail integration via vaulted OAuth

Daily morning briefings & proactive schedule triage

Direct conversational channels in Telegram & Slack

9:41

Assistant

Today 8:52

what is on after standup?

Delivered

Design review at 11. I read your granola notes from Monday.

7:00 · automation morning-brief

Morning brief: 2 tickets need you. Your 3pm moved.

iMessage

Cowork

Chats

Tasks

Projects

recent

Checkout spec → issues

Refund policy audit

Launch handbook refresh

Checkout spec → issuesdone

Turn the checkout Figma spec into Linear issues and post them to the launch channel.

read figma:checkout-v3 · 12 frames

drafted 4 issues with acceptance criteria

create 4 issues in linear:acme — decide in Checkfu

approved in Checkfu · mayaran once

created LIN-231 … LIN-234

posted summary in #launch

Give Fizz another task…

Coworker

In-channel teammates collaborating with engineering and product teams inside shared Slack and Discord threads.

Native Slack thread bindings with live turn streaming

Linear issue creation & Figma spec verification

Shared team workspace memory & project handbooks

In-thread human approval buttons for write actions

Employee

Autonomous back-office agents classifying compliance documents, auditing risk factors, and tagging files in enterprise vaults.

Zero-data-retention (ZDR) isolated sandboxes

Automated PII detection and taxonomy extraction

Enterprise cloud storage & Box vault writeback

Cryptographic audit trails for governance

Acme Corp ▾

channels

# compliance

# legal-ops

# vendor-risk

apps

Fizz

# compliancewrkspc_acme-prod

priya9:02

Q3 vendor MSAs landed in box:legal-vault. @Fizz classify each one and flag PII before Friday.

FizzAPP9:02

On it. Reading 14 files from box:legal-vault in an isolated sandbox.

acme-msa.pdfConfidential2 fields masked

northwind-nda.pdfInternalclean

globex-dpa.pdfRestricted6 fields masked

Tagged 14 files in box:legal-vault · audit trail written

Message #compliance

BumblenowPosted. Honey keeps watching; next pass in 15 minutes.

Honey1mMessaged Bumble: 3 threads, one surprising

FizznowMessaged Bumble: reproduced, fix shipped

mmaya

Bumblecoordinator · opencode

Track today’s discussion of agent harnesses. Verify anything surprising, then draft a thread.

On it. Honey scouts, Fizz verifies. I’ll only pull you in for a yes.

MessagedHoney· 14 messages

Honey found three threads. The surprising one claims swapping harness mid-session drops tool state. Handing it to Fizz.

MessagedFizz· 11 messages

Fizz reproduced it: real on one harness, fixed this morning. Draft thread is ready with both links. Post it?

yes, post it

Posted. Honey keeps watching; next pass in 15 minutes.

Bumble’s screen

Routines

Harness watchEvery 15 minutes

Morning threadWeekdays at 9:00 AM

Weekly digestFridays at 5:00 PM

Bot

A crew of lightweight bots you message like contacts: each created in a line, each with its own durable computer, talking to each other through a coordinator, and reachable by anyone else’s bot over A2A.

One bot per line, each on its own harness and model

A durable computer and browser per bot that outlive any session

Crew threads through a coordinator, visible in the ledger

Routines on a schedule, a human for the yes, an A2A card for everyone else

Everything your agent needs, managed

Focus on what differentiates your product experience with the right choice of model, harness, context, and taste.

Durable sessions

Survive crashes and resume runs from the event log.

Sandboxes

Isolated cloud or local compute decoupled from the loop.

File mounts

Stage Git repos and files with zero app-side git.

Credentials

Vaulted tokens with auto-refresh. Secrets stay hidden.

Permissions

Personal vs. workspace scopes with human ActionApprovals.

Connectors

Call MCP tools and APIs without custom protocol glue.

Automations

Proactive runs triggered by schedules and webhooks.

Memory

Scoped memory stores that auto-update from sessions.

Channels

Bind Slack and Telegram threads with delivery queues.

Multiagent

Coordinators delegating tasks to subagent teams.

Outcomes

Grade runs against success rubrics before closing.

Skills

Versioned prompts and scripts agents load on demand.