Ticket moved to ReadyLinearGitHubVercelSlack
Delivers: A tested pull request with a preview link, ready for review
Software engineering
Feature Builder
Builds each ready ticket into a working feature, with tests, in its own sandbox.
Managed AI agents for every team
Recursion Managed Agents puts AI agents to work inside the apps your business already uses: Salesforce, Zendesk, NetSuite, Slack, Jira and 100+ more. They start on a schedule, a message or an API call, finish the job, and get better every run.
Recursion Managed Agents· Sessions
Recursion Managed Agents· Sessionsincident INC-4821
INC-4821 · SEV-2 · checkout-api · 5xx at 6× baseline since 14:07
Incident: checkout is failing for some customers after the 14:05 deploy
Find the root cause, stop the errors, ship a tested fix, count the customer impact and keep the status page current.
running
8 agents · 4 models
outcome satisfied 5/5 · mitigated in 26 min · fix in review
- PagerDuty0
- Datadog0
- GitHub0
- LaunchDarkly0
- Stripe0
- Zendesk0
- Statuspage0
- Slack0
coordinator · PagerDuty INC-4821 acknowledged · checkout-api 5xx at 6× baseline since 14:07 · plan: 5 checks
Checkout errors spike after a deploy.
PagerDuty pages the on-call: checkout is failing at six times the normal rate since 14:07, two minutes after the 14:05 deploy. The coordinator reads the alert, the error graphs and the deploy, then plans five checks.
Five specialists investigate in parallel.
Logs, the deploy diff, feature flags, a reproduction and customer impact, each on the model it is best at, across Datadog, GitHub, LaunchDarkly, Stripe and Zendesk. The coordinator waits for the on-call’s OK to turn the flag off, then a sixth specialist writes the fix.
That was one of 144 sessions this week.
Engineering and research agents on one board: done, running or waiting. Brighter means better graded.
Every finished session teaches the agent something.
Each session leaves a short report. Once five are ready, a memory run verifies what they teach and updates the agent’s memory, and the next session reads it before it acts.
One goal in. A fleet of specialists out.
Recursion Managed Agents runs persistent, long-horizon agents as one fleet across your repos, CI, clouds and data. Hand them a goal, not a script: fix this regression, find the fastest config, keep every dependency current. Sandboxes, credentials, subagents, memory and grading are all handled for you.
Works across your stack
100+ integrations across code, CI, incidents, data and docs. Agents take real actions, limited to the tools you allow.
Runs on its own, for days or weeks
Starts on a failing build, an alert, a merged pull request, a schedule or an API call. Sessions pause on dependencies and pick up where they left off.
Multi-agent by design
A coordinator plans the work and spawns specialists in parallel. Big jobs get more agents, not longer prompts.
Every outcome graded
A separate grader checks each session against your rubric and sends it back until every criterion passes. Every finished session feeds the agent’s memory.
Know exactly how your agents are performing
A coordinator breaks down each request and sends the right checks to specialist agents with the models and apps they need. The specialists gather the evidence, and the coordinator brings it together into a recommendation.
A separate grader evaluates the result against your rubric, criterion by criterion, using the evidence gathered along the way.
Set the standard, measure every outcome, and manage agent quality at scale.
Recursion Managed Agents· Sessions
Recursion Managed Agents· Sessionscoordinator · 6 specialists · 1 grader
Inference coordinatorClaude
JiraENG-2291
Cut the ranking model’s p95 latency by 30% without losing accuracy.
planreading ENG-2291 and the code
Profiler
Quantization
Batching
Distillation
Load test
Accuracy guard
recommendation
Ship int8 with a 4 ms batch window: p95 −38%, accuracy −0.1.
ProfilerGPTDatadog
not started
QuantizationClaudeGitHubSandbox
not started
BatchingOpen weightsGitHubSandbox
not started
DistillationGeminiGoogle Cloud
not started
Load testGPTSandboxGrafana
not started
Accuracy guardClaudeGoogle BigQuerySandbox
not started
Outcome graderClaude
your rubricwaiting for a recommendation
Bottleneck found in production
evidence: Datadog traces, 10k requests
Change is tested and reviewed
evidence: PR #5120, CI green
No new hardware
evidence: same GPU pool, 2.3× throughput
p95 down at least 30%
evidence: load test at 2× peak traffic
Accuracy within 0.2 points
evidence: 50k held-out queries, 3 seeds
satisfied 5/5ship to a 5% canary
See your entire agent fleet at a glance
See every agent at work in one place, from flaky tests and incidents to eval reruns and experiment sweeps. Know what’s done, what’s in progress, and how well each agent is performing.
Get a live view of your operations and the quality of every run.
Recursion Managed Agents· Sessions
Recursion Managed Agents· Sessions144 sessions this week · 127 completed · 10 running
- Incident Investigation
- Flaky Test Fixer
- Eval Regression Watch
- Experiment Sweep
- brighter = better graded
Assistants wait for an engineer. The backlog doesn’t.
Chat and coding assistantsworking hours
A person starts and steers every session
Recursion Managed Agentsall 168 hours
SchedulesEventsAPI calls
Engineering work repeats
Red builds, flaky tests, dependency bumps, an eval rerun on every checkpoint, a sweep that needs watching overnight. Most of it comes back on a schedule or lands in a queue, follows the same steps each time, and there’s more of it than your team can get to.
Assistants wait for a person
Your engineers use coding assistants all day, and they ship faster for it. But every session starts with someone at the keyboard: someone has to notice the red build, start the work, check the result and push it where it belongs. When they log off, the queue keeps growing.
Recursion Managed Agents runs it on its own
Its agents start on a failing build, an alert, a new checkpoint or a schedule, run hundreds of sessions in parallel when the queue is long, and hand back pull requests and results. Nobody has to press go.
Work that runs in the background
Each agent has one job, the tools it may use and what starts it: a failing build, an alert, a new checkpoint or a schedule. No one has to open a chat and ask. These are a few of the jobs engineering and research teams run on it; any work that keeps coming back can be one.
Works across your engineering stack
Connect 100+ tools across code, CI, incidents, data and docs, then choose exactly which tools each agent may call: read only, changes data, or destructive. Credentials stay in a vault, and every tool call is recorded in the session transcript.
Recursion Managed Agents· Integrations
GitHub
GitHub
Native integration · vaulted credential
Tools this agent can call4 of 5 on
- get_pull_requestread only
- list_check_runsread only
- create_branchchanges data
- create_pull_requestchanges data
- merge_pull_requestdestructive
Credentials stay in the vault. The agent sees only the tools you turn on, and every call lands in the session transcript.
Native integrations
Slack, GitHub, Jira, Confluence, Google Cloud Logging, Loom and LaunchDarkly are built into Recursion Managed Agents. Start an agent from a Slack message and get the result back where your team works.
100+ built-in connectors
Datadog, PagerDuty, Sentry, Snowflake, Databricks and the rest of the catalog. Connect once, then turn tools on per agent.
Your own MCP servers
Point an agent at any MCP server, internal or third party, and set the same per-tool permissions.
Browser tasks, handled on a schedule
Some work only happens in a browser: vendor portals behind a login, sites with no API, forms that have to be filled in by hand. With computer use, the agent gets its own browser to sign in, search, fill in forms and download files. Put it on a schedule and it does the rounds every week on its own. Every click is saved as a screenshot you can replay.
Recursion Managed Agents· Sessions
Recursion Managed Agents· SessionsRFP Sweep · scheduled run · computer use
New tab
New tabSearch or type a URLAgent driving
Agent
Computer · every action saved with a screenshot0 frames
Live
RFP Sweepsession 7318running
every Monday 07:00computer useSalesforceSalesforce
Agent narration
Monday 07:00 run. Checking 6 bid portals for new RFPs that fit our fleet software. None has an API, so I’ll sign in to each one.
Runsevery Monday 07:00
Sep 21
0 new
Sep 28
1 new
now · Oct 5
running
next · Oct 12
07:00
Your agent fleet, in your pocket
Your agents keep working after you log off. Check on them from your iPhone or iPad: see every session on one board, follow a run as it happens, and open the reports and files they made. Widgets on your Home Screen and Lock Screen show what’s running at a glance.
Fine-tune specialist models and add them to the fleet
Memory makes each run better. Fine-tuning makes the model itself better. When a job runs at volume, Recursion Managed Agents turns its graded runs into environments and evals, trains a specialist with reinforcement learning, and adds it to your fleet when it beats the model the agent runs today.
- 1Graded runs12,480 graded runs
Every session scored against your rubric
- 2Environments
- Refund above policy limit
- Duplicate charge dispute
- Order number missing
- +1,237 more
Real tickets become repeatable scenarios
- 3Evals312 held-out tickets
- Right resolution
- Policy followed
- No invented facts
- Escalates when needed
Held-out checks on every criterion
- 4RL trainingreward0.61 → 0.88
Rewarded for passing your rubric
- 5In the fleetSupport TriageClaude Opus 4.8Support specialist v384% resolved · $32 per 1k tickets
Promoted when it beats the model it replaces
The specialist always beats the generalist
On the job it was trained for, the specialist resolves more tickets than frontier models, invents less and answers faster, at a fraction of the cost.
- lower cost per task with fine-tuned models
- 5–10×
- of leading US AI labs trust Labelbox Horizon's RL environments
- 90%+
Support Triage after fine-tuning: more tickets resolved, fewer hallucinations, faster responses and lower inference cost than frontier models
Resolution rate
Support specialist v3
84%GPT-5.5
76%Claude Opus 4.8
73%Reduction in hallucinations
Support specialist v3
72%GPT-5.5
46%Claude Opus 4.8
41%Cost per 1,000 support tickets
Support specialist v3
$32GPT-5.5
$158Claude Opus 4.8
$176Time to first token
Support specialist v3
0.42sGPT-5.5
1.10sClaude Opus 4.8
1.28sManaged agents, without the lock-in
Agent platforms from the AI labs run only their own models. Recursion Managed Agents runs any model on any job, including your own checkpoints, works across the tools your engineers already use, and turns every graded run into a better specialist.
- Built in
- Partly, or you build it
- Not offered
Models
Recursion Managed AgentsBuilt in: Any model, per agentFrontier, open-weight or your own fine-tuned modelsOther managed agentsNot offered: One lab’s modelsLocked to the providerMulti-agent
Recursion Managed AgentsBuilt in: A team of specialistsEach with its own model, tools and accessOther managed agentsPartly, or you build it: SubagentsAll on the same lab’s modelsIntegrations
Recursion Managed AgentsBuilt in: 100+ integrations, built inPer-tool permissions, credentials in a vaultOther managed agentsPartly, or you build it: Bring your ownYou connect and maintain MCP serversGrading
Recursion Managed AgentsBuilt in: Every run gradedA separate grader, revising until it passesOther managed agentsPartly, or you build it: Not always built inOften yours to buildMemory
Recursion Managed AgentsBuilt in: Learns automaticallyEvery run makes the next one betterOther managed agentsPartly, or you build it: Varies by providerMemory stores or session historyTraining
Recursion Managed AgentsBuilt in: Specialist models from your runsGraded runs become evals and RL trainingOther managed agentsPartly, or you build it: Separate fine-tuning, where offeredYou bring the data and graders
Other managed agents: the leading AI labs' agent platforms, based on their public documentation as of September 2026.
Real access to your systems. Never the keys.
Every session runs in an isolated sandbox with only the tools and credentials you grant it. Recursion Managed Agents attaches each credential to the approved call, so the model never sees it, and keeps it to the hosts you allow. Every model turn and tool call lands in an audit-ready transcript, backed by Labelbox's enterprise security and compliance program.
- SOC 2 Type II
- GDPR
- CCPA
- NIST 800-171
- PCI DSS
Put your first agent to work
Start with the job your engineers repeat every week: a red build, a flaky suite, an eval rerun. Connect the repos and tools it touches, describe what done looks like and pick what starts it. From then on, it gets done before anyone thinks to ask.