Operations6 min read
The approval queue is the product
Teams that ship fastest treat the queue like a code review lane: small diffs, fast feedback, and a rule written into memory every time they reject something.
42 agents online · 8 keys rotating
The Agent Brain of Your AI Stack
From prompts to production fleets of agents — direct a crew that perceives, plans, acts, and remembers. All in one control room.
1,284
Signals
17
Plans
42
Runs
128k
Memory
Signals
Perception
1,284events / hr
Salesforce · new deal · Acme Corp — $42k
Slack · #alerts · workflow-14 failed retry
Webhook · stripe.invoice.paid · $1,240
Notion · doc updated · Q4 pricing
Every webhook, inbox, DB change and API poll — one stream.
120+ connectorsat-least-once deliveryPII redactionreplayable log
Plans
Reasoning
17drafts pending
Auto-KYB triage · 4 steps · confidence 92%
Cost-cutting agent · 3 steps · savings est. $3.2k/mo
Slack failure alerts · 2 steps · draft
EU expansion review · 6 steps · needs approval
Multi-step plans, drafted by agents, approved by you.
chain-of-thought traceshuman approval gatescost estimatesbranch & merge
Runs
Agency
42live now
run-8f2a · KYB check · Acme Corp · 34s
run-8f2b · pricing scrape · 12 sites · 2.4m
run-8f2c · lead enrichment · 240 rows · queued
run-8f2d · Slack digest · daily · 4s
Agents that actually do the work — safely, in parallel.
sandboxed toolsretry & rollbackcost caps per runaudit trail
Memory
History
128kdecisions logged
learned: reject KYB when UBO < 20% + high-risk geo
learned: retry Slack once, then escalate to email
outcome: 92% CSAT on auto-support (last 30d)
outcome: $12.4k saved via cost-cutting agent (30d)
Every decision, outcome, and lesson — searchable forever.
vector + relationalSOC2 retentionoutcome scoringpolicy learning
Suggested for you
predicted from recent activityLive telemetry
The room, right now
key rotation · 8/8 healthy
event streamtail -f room.log
- just nowwebhookStripe → invoice.paid · $1,240 · Acme Corp
- 2sdecisionrun-8f2a · KYB verdict: PASS (confidence 0.94)
- 6splanDraft: renewals-at-risk — 4 steps, ~$0.02 / run
- 11slearnLogged outcome: onboarding SLA hit for 43/45 accounts
- 18scdcSalesforce · new opportunity · Northwind — $84k
- 27sjobrun-8f2b · scraped 12 competitor pricing pages in 2.4m
- 34sgateApproval requested · EU expansion review · 6 steps
- 41spolicyPolicy updated: retry Slack once → escalate to email
P95 decision latency
1.8s-0.4s wow
Spend saved (30d)
$12.4k+3.1k vs last mo
Human time reclaimed
284 h7 FTE-days
Approval hit rate
97.2%on 1,240 gates
Lifecycle
One loop, four moves, running while you sleep
Every workflow in ReplitAgents walks the same loop. It is boring on purpose — boring is what makes it auditable.
01 · Observe
The room subscribes to every surface you already pay for — CRM, ledger, inbox, warehouse, on-call.
- Normalised event envelope
- Dedup + ordering guarantees
- Redaction before inference
median 240ms ingest → decision
02 · Deliberate
An Architect agent drafts the plan, prices it, and shows its reasoning before a single tool fires.
- Cost + latency estimate per step
- Alternate branches scored
- Policy check against your rules
avg 4.2 steps per approved plan
03 · Act
Operators execute in a sandbox with scoped credentials, hard cost caps, and rollback on every write.
- Parallel fan-out with backpressure
- Idempotency keys on writes
- Circuit breakers per tool
99.2% runs complete without human touch
04 · Learn
Outcomes are scored and written back, so the next plan starts smarter than the last one did.
- Outcome → policy distillation
- Regression replay on old runs
- Searchable decision ledger
31% fewer human escalations by week 6
The crew
Meet the agents on shift
S
Sentinel
Compliance & KYB
Watches every new counterparty. Flags UBOs, sanctions, adverse media in seconds.
OpenCorporatesOFACChainalysisSardine
Runs (30d)
12,480
Confidence
0.94
A
Architect
Workflow designer
Turns a one-line goal into an approvable multi-step plan with cost estimates.
PlaybooksDAG editorCost modelSimulator
Runs (30d)
3,120
Confidence
0.91
O
Operator
Execution engine
Runs plans in parallel with sandboxed tools, retries, and rollback on failure.
HTTPSQLBrowserShell (jailed)
Runs (30d)
184,204
Confidence
0.97
A
Archivist
Memory & learning
Scores outcomes, distills lessons, and rewrites the crew's playbooks nightly.
Vector storeOutcome tagsPolicy graph
Runs (30d)
n/a
Confidence
0.93
S
Scribe
Docs & comms
Drafts every summary, changelog, and stakeholder update. Signs off on tone.
NotionSlackEmailPDF
Runs (30d)
9,840
Confidence
0.89
N
Negotiator
Cost-cutting
Audits SaaS spend, benchmarks contracts, and drafts renegotiation outreach.
Vendr dataG2EmailCalendar
Runs (30d)
412
Confidence
0.88
Capabilities
Everything you'd build in-house, minus the six-month build.
Visual workflow designer
Draw agent graphs on an infinite canvas. Nodes are tools, plans, or humans.
Conversational console
Delegate in plain English. The crew asks clarifying questions, not vice-versa.
Branch, simulate, ship
Fork a workflow, replay last week's traffic, diff outcomes before rollout.
Human-in-the-loop gates
Approve high-stakes actions from Slack, email, or the room itself.
Sandboxed tool execution
Every tool call runs isolated with per-tenant creds, quotas, and audit trail.
Outcome-based scoring
Agents don't just log — they learn from what worked, and rewrite plans.
8-key rotating inference
Groq keys round-robin with cooldown. Rate limits are the platform's problem.
Composable primitives
Signals, Plans, Runs, Memory — mix, nest, and reuse across the org.
Inference fabric
Eight keys, one brain, zero rate-limit walls
Live pool state, polled every 10s from GET /api/health. Requests round-robin across the Groq keys. A 429 penalises the key, a 5xx cools it down, and traffic reroutes mid-flight — the caller never sees a retry.
GROQ_API_KEY_1
ready
0 fails · healthy
GROQ_API_KEY_2
ready
0 fails · healthy
GROQ_API_KEY_3
ready
0 fails · healthy
GROQ_API_KEY_4
ready
0 fails · healthy
GROQ_API_KEY_5
ready
0 fails · healthy
GROQ_API_KEY_6
ready
0 fails · healthy
GROQ_API_KEY_7
ready
0 fails · healthy
GROQ_API_KEY_8
ready
0 fails · healthy
Sub-second first token
p50 340ms to first token across the pool, measured from the edge worker.
Deterministic failover
Round-robin with exponential penalty; a bad key is skipped for 60s, then probed.
Keys never leave the server
Rotation happens inside server functions — the browser only ever sees tokens.
Integrations
Wired into your stack in minutes
Every connector ships with typed operations, an auth posture, a rate-limit budget the scheduler respects, real setup steps, the events it emits, and the mistake teams make first. Click one to see exactly what an agent can and cannot do with it.
Postgres
Data- Auth
- Conn string + mTLS
- Scopes
- read-only role
- Operations
- query, stream_changes, explain, schema_introspect
- p50 latency
- 38 ms
- Budget
- 200 q/s
- Cost
- $0 — your database
Setup
- 1Create a role: CREATE ROLE agent LOGIN PASSWORD '…' NOINHERIT;
- 2GRANT SELECT on only the schemas the room needs.
- 3Paste the replica connection string; the gateway verifies TLS and introspects the schema in one pass.
Signals emitted
row_insertedrow_updatedlogical_slot_lag
Default guardrail
Statement timeout pinned to 15s; agents get a read replica by default.
Most common mistake
Pointing an agent at the primary. Always use a replica — a runaway aggregate should never touch write traffic.
room.connect("postgres", {
auth: "Conn string + mTLS",
allow: ["query", "stream_changes", "explain", "schema_introspect"],
budget: "200 q/s",
gate: ["write", "spend"],
})Auth handled for you
OAuth refresh, key rotation, STS assume-role, and mTLS all live in the gateway — never in your workflow code.
Rate limits are budgets
The scheduler reads each connector's budget and paces agent fan-out, so you never burn a partner quota.
Every call is replayable
Request, response, latency, and the deciding prompt are stored on the run trace for 30 days.
Bring your own tool
An OpenAPI spec or an MCP server endpoint becomes a first-class, typed agent tool with zero adapter code.
Playbooks
Battle-tested workflows, one click to fork
Zero-touch KYB
New signup → sanctions + UBO + adverse media → verdict in <2 min. Escalate the gray zone.
1. Signup webhook2. Vendor fanout3. Score & merge4. Post to Salesforce
SaaS spend hunter
Reads Ramp + SSO logs, finds dead seats, drafts renegotiation emails, books calls.
1. Pull usage2. Cluster wasteage3. Draft outreach4. Human approve
Renewal defender
Watches product usage + support tone, warns 45 days before churn, briefs the CSM.
1. Signal blend2. Risk score3. CSM brief4. Save-play
Incident narrator
PagerDuty fires → agent scrapes logs, drafts customer message, files post-mortem.
1. Trigger2. Log dive3. Draft note4. Post-mortem
Template gallery
Start from a room someone already proved out
Every template ships with its tool contracts, guardrails, and the cost cap the original author used.
1,240 installs
Zero-touch KYB
Registry lookup → UBO graph → sanctions screen → risk memo
Fork template980 installs
Churn early-warning
Usage drop → support sentiment → CSM brief in Slack
Fork template2,110 installs
SaaS spend hunter
SSO logs → seat audit → renewal calendar → cancel draft
Fork template3,480 installs
Inbound SDR
Form fill → enrich → score → route → personalised reply
Fork template760 installs
Invoice reconciler
Ledger diff → vendor match → exception queue → journal entry
Fork template1,530 installs
On-call scribe
Pager event → timeline build → post-mortem draft → Linear issue
Fork template890 installs
Contract redliner
Upload → clause diff vs. playbook → risk flags → counsel summary
Fork template1,870 installs
Competitive radar
Pricing pages → changelogs → job posts → weekly brief
Fork template640 installs
Renewal defender
Usage forecast → discount model → outreach sequence → close loop
Fork template184k
agent runs / month
1.8s
P95 decision latency
8×
Groq keys, hot-swapped
97.2%
approval accuracy
120+
native connectors
$12.4k
avg saved / room / mo
The math
What the room gives back, per month
Modelled on a 60-person company at a $78/hr fully-loaded rate. Your mileage will differ; the shape rarely does.
| Task | Human | Agent | Volume | Recovered |
|---|---|---|---|---|
| KYB / vendor onboarding review | 45 min | 38 sec | 620 / mo | $18,400 |
| Inbound lead enrichment + routing | 12 min | 6 sec | 3,400 / mo | $27,900 |
| SaaS renewal & usage audit | 6 hrs | 4 min | 40 / mo | $11,200 |
| Tier-1 support triage & reply | 9 min | 11 sec | 7,800 / mo | $46,300 |
| Invoice exception reconciliation | 22 min | 19 sec | 1,150 / mo | $16,700 |
| Incident post-mortem drafting | 3 hrs | 90 sec | 28 / mo | $8,900 |
1,940 hrs
human time returned / mo
$129.4k
modelled monthly recovery
11.4×
return against platform cost
Compared
Why teams switch to ReplitAgents
| Capability | ReplitAgents | Generic chatbot | In-house scripts |
|---|---|---|---|
| Infinite canvas of agents | |||
| Signals / Plans / Runs / Memory model | |||
| 8-key rotating inference | |||
| Human-in-the-loop gates | |||
| Sandboxed tool exec + audit | |||
| Outcome-based memory | |||
| Time-to-first-agent < 5 min | |||
| Cost caps per run |
Straight answers
The six things every buyer asks us
No hedging. These are the answers we give in the second call, written down so you can skip it.
Guardrails
“Agents hallucinate — I can't put that near production.”
Nothing writes without a typed tool contract and a policy check. Anything below your confidence floor becomes a draft in the approval queue instead of an action.
Coexistence
“We already have automations in Zapier / n8n.”
Keep them. Signals ingests their webhooks, and Runs can call them as tools. ReplitAgents is the layer that decides which of them to fire, and why.
Deployment
“Our data can't leave our VPC.”
Single-tenant deploy into your VPC, BYO KMS, and BYO inference endpoint. The control plane holds metadata only.
Accountability
“Who is on the hook when an agent gets it wrong?”
Every run carries an immutable trace: the signal, the plan, the approver, the tool calls, the diff. Roll back one run or an entire day.
Cost
“LLM spend is unpredictable.”
Cost caps at workspace, plan, and run level, with hard stops rather than alerts. You see the estimate before you approve.
Adoption
“Our team won't learn another tool.”
They won't. Most operators live in Slack — approvals, digests, and escalations arrive there. The control room is for when you want the whole picture.
Industries
Same room, tuned for your world
Fintech & banking
42h → 90s KYBKYB, AML, transaction monitoring, dispute triage — all with audit trails your regulator will love.
SentinelOperator
Healthcare & life sciences
63% first-pass approvalsPrior-auth drafts, claim scrubbing, referral routing. HIPAA-aligned tenants, PHI redaction by default.
ScribeArchivist
E-commerce & retail
$4.20 CAC saved / orderReturns triage, pricing scrapes, catalog enrichment, ad-copy A/B — 24/7 across 40+ storefronts.
OperatorScribe
Legal & compliance
6h → 22m reviewRedline diffs, precedent search, matter intake, docket watch. Every citation grounded, no hallucinated cases.
ArchivistScribe
B2B SaaS ops
37% saves liftRenewal defense, expansion signals, churn briefs, MRR reconciliation. Straight into CRM.
ArchitectNegotiator
Supply chain & ops
11d faster claimsPO exceptions, carrier ETAs, damage claims, supplier scorecards. Alerts before the SLA slips.
SentinelOperator
Architecture
One loop, four organs
Signals
layer 1 / 4
Webhooks · CDC · polls · inbox
t+0ms
Plans
layer 2 / 4
LLM plans · gates · simulator
t+120ms
Runs
layer 3 / 4
Sandboxed tools · retries · caps
t+240ms
Memory
layer 4 / 4
Outcomes · vectors · policies
t+360ms
event.received → plan.drafted → gate.approved → run.dispatched → tool.invoked → outcome.scored → policy.updated — closed loop in < 2s median.
Observability
If you can't see it, you shouldn't automate it
p95 2.4s
Run traces
Waterfall of every tool call with tokens, latency, and cost attributed per step.
6 monitors
Drift watch
Alerts when a plan's approval rate or outcome score moves outside its baseline band.
$0.0041 / run
Token ledger
Spend by agent, workflow, tenant, and key — exportable to your warehouse nightly.
418 blocks / mo
Policy hits
Every blocked action, who set the rule, and what the agent tried to do instead.
unlimited
Replay
Re-run any historical signal against a new plan version to see what would change.
median 4m
Human loop stats
Approval latency by reviewer, so you can find the bottleneck that is a person.
Developer surface
One SDK, every language your stack speaks
POST /v1/plans
import { room } from "@replitagents/sdk";
const plan = await room.plan({
goal: "Onboard Acme Corp with a full KYB pass",
approvalChannel: "slack:#compliance",
});
const run = await plan.execute({ costCap: 0.25 });
console.log(run.outcome); // { verdict: "PASS", confidence: 0.94 }Console grammar
Eight verbs run the entire platform
Plain English works everywhere. Slash commands are for when you already know exactly what you want.
/plan
Draft a multi-step workflow from a goal, priced and scored before anything runs.
/plan cut our Datadog bill 30% without losing p1 coverage
/run
Execute an approved plan now, or dry-run it against last week's signals.
/run kyb-triage --dry --since 7d
/watch
Turn any query into a standing signal that wakes the room when it changes.
/watch stripe disputes > 3 in 24h
/remember
Write a durable fact or policy into memory that every future plan must respect.
/remember never auto-approve vendors in high-risk geos
/why
Ask any run to justify itself — full trace, tool calls, and the rule that applied.
/why run-8f2a rejected Acme Corp
/undo
Roll back a run, a batch, or an entire day of agent writes.
/undo --since 09:00 --agent operator
/cost
Break spend down by agent, workflow, or tenant for any window.
/cost --by workflow --last 30d
/handoff
Escalate to a human with the full context attached, in Slack or email.
/handoff @maya run-8f2c needs pricing sign-off
Benchmarks
Measured, not marketed
| Task | ReplitAgents | Chatbot | Human baseline | Advantage |
|---|---|---|---|---|
| KYB decision (mid-risk) | 1.8s | 42s | 42h | 23× faster |
| Refund triage | 0.9s | 6s | 8m | 533× faster |
| Renewal risk brief | 3.2s | n/a | 35m | 656× |
| Incident post-mortem draft | 14s | 1m | 3h | 771× |
| SaaS spend audit (per vendor) | 22s | n/a | 40m | 109× |
Case studies
Rooms that closed the loop
Northwindcase study
$1.2M in fraud stopped in Q3 alone
Sentinel + Operator run in-line on every wire above $10k. False positive rate dropped from 14% to 3.1% in 60 days.
-83%
Fraud losses
-71%
Analyst hours
0
SLA breach
Helianthuscase study
Renewal saves lifted 37% quarter over quarter
Architect briefs every CSM 45 days out with usage decline, sentiment, and a ready save-play. Approvals happen from Slack.
+37%
Save rate
-4h/day
CSM prep
118%
NRR
Vermilion Healthcase study
Prior-auth turnaround from 6 days to 11 hours
Scribe drafts every submission grounded in payer policy Archivist keeps up-to-date. Human clinicians sign off in one click.
-92%
TAT
-44%
Denials
63%
First-pass
Trust
Built for the room your legal team will sign off on
SOC 2 Type II
audited annually
ISO 27001
in progress · Q1
HIPAA-ready
BAA on request
GDPR + DPA
EU tenants available
BYO KMS
AWS KMS · GCP KMS
VPC deploy
single-tenant option
- Per-tenant tool credentials, never shared across rooms
- Full audit log — every prompt, tool call, and human approval — exportable
- PII detection & redaction on every signal before it hits an LLM
- Cost caps + circuit breakers at plan, run, and workspace level
- 8-key rotating inference with per-key cooldown; no vendor lock-in
- SSO (SAML + OIDC), SCIM provisioning, and granular RBAC on every surface
Deployment
Run it where your data already lives
Cloud
Teams who want it running this afternoon
- Multi-tenant, EU or US region
- Managed 8-key inference pool
- 99.9% SLA on the control plane
- Onboard in <10 minutes
Private region
Regulated teams with residency rules
- Dedicated database + object store
- Your region, your retention window
- 99.95% SLA, named support
- SCIM + SAML included
In your VPC
Data that legally cannot leave
- Single-tenant Helm / Terraform
- BYO KMS and BYO inference endpoint
- Control plane sees metadata only
- Air-gapped mode on request
Getting started
From signed-up to autonomous in six weeks
Day 0
Connect two sources
OAuth your CRM and your ledger. Signals starts streaming in under ten minutes — read-only, nothing fires yet.
Day 1
Watch the shadow runs
Agents draft plans against live signals but execute nothing. You review the diff of what would have happened.
Day 3
Approve your first playbook
Promote one shadow workflow to live with a human gate on every write and a $50/day cost cap.
Week 2
Drop the gate on the boring 80%
Runs above 95% historical accuracy go fully autonomous. The rest keep the approval step.
Week 6
Memory starts paying rent
Outcome scoring has tuned thresholds; escalations fall, and the room proposes its own next playbook.
Roadmap
What's on the workbench
Shipped
- 8-key Groq rotation
- Infinite canvas v2
- Slack approval gates
- Cost caps per run
- Outcome-scored memory (beta)
Now
- Multi-model routing (Groq / Anthropic / OpenAI)
- Native MCP tool registry
- Workflow diff & replay
- VPC deploy self-serve
Next
- Voice-first console
- Real-time collaborative canvas
- Auto-generated post-mortems
- Room templates marketplace
Exploring
- Federated memory across rooms
- On-device Signals for edge stacks
- Agent-to-agent negotiation protocol
Principles
How we decide what to build
Six rules we argue about in every design review. When something feels off in the product, it's usually one of these being broken.
01
Delegation beats navigation
If a user has to find a menu, we failed. State the outcome; the room finds the path.
02
Show the reasoning, always
An agent that can't explain itself is a liability. Traces are a first-class product surface, not a debug tab.
03
Autonomy is earned per task
Trust is granted workflow by workflow, backed by measured accuracy — never toggled globally.
04
Reversibility over speed
Any write an agent makes, a human can undo in one click, up to and including a whole day of runs.
05
The interface adapts to the operator
A finance lead and an SRE open the same URL and see different rooms. Same data, different affordances.
06
Boring infrastructure, ambitious surface
Round-robin keys, idempotent writes, append-only logs. Save the novelty for the experience.
The humans
Operators building for operators
AN
Ari Nakamura
CEO · ex-Stripe payments
JO
Jules Okafor
CTO · ex-Anthropic infra
MS
Maya Solberg
Head of Product · ex-Linear
DR
Devansh Rao
Head of Research · ex-DeepMind
EP
Elena Petrov
Head of Design · ex-Figma
KA
Kwame Asare
Head of Trust · ex-Cloudflare
Featured in
"the first agent platform that admits agents need a control room" — Latent Space
TechCrunchThe InformationBloombergSiftedThe VergeWiredLatent SpaceAxios
Product Hunt · #1 of the day
GitHub Trending · 3 weeks
a16z · Top 50 AI apps 2026
Y Combinator W25 alumni
Field notes
What we've learned running rooms in production
Engineering9 min read
Why we round-robin eight keys instead of buying one big quota
Rate limits are bursty and per-key. Eight small keys with penalty-aware routing beat one large key on both p95 latency and cost per successful call.
Practice5 min read
Confidence thresholds are a business decision, not an ML one
Set the floor by asking what a false positive costs versus a missed one. We now ship a worksheet for this in every onboarding.
Research11 min read
Memory is worthless without outcome labels
Storing transcripts is easy. Storing what happened next — and whether it was good — is the part that makes the next plan better.
Glossary
Speak room, fluently
Room
Your workspace — an infinite control room hosting one crew of agents and their memory.
Signal
An inbound event: webhook, DB change, poll, inbox — anything that wakes an agent.
Plan
A multi-step, cost-estimated, human-approvable graph of tool calls.
Run
A single execution of a plan with sandboxed tools, retries, and a full trace.
Gate
A human approval checkpoint — Slack, email, or the room UI.
Playbook
A parameterized template of a plan you can fork and re-tune.
Outcome
The scored result of a run: hit, miss, needs-review — the primary training signal.
Cooldown
Time a Groq API key sits out after a 429/5xx before rotating back in.
Guides & resources
Read the room
Working guides, not marketing posts — each carries prerequisites, the exact steps, a copyable snippet, the pitfalls that bite teams first, a go-live checklist, and the one thing worth remembering.
Go from an empty workspace to a live room that ingests a webhook, plans a response, executes two tool calls, and writes what it learned to Memory.
Prerequisites · A webhook source and one connector (Slack is easiest).
- 1Create a room and name the outcome, not the task — 'disputes answered', not 'read disputes'.
- 2Point a webhook at the room's ingest URL; the raw payload is stored and becomes replayable immediately.
- 3Open the console and describe the outcome in plain English. The compiler emits a 4–7 step spec.
- 4Dry-run the spec in the Studio: every tool call is simulated and the guardrails are checked, nothing is written.
- 5Promote to live with a gate on any step that spends money or emails a human.
curl -X POST https://api.replitagents.com/v1/plans \
-H "Authorization: Bearer $RA_KEY" \
-d '{"goal":"Answer Stripe disputes with evidence","gate":["refund"]}'Pitfalls
- ·Naming the room after a tool ('Slack bot') instead of an outcome — the compiler then plans around the tool, not the goal.
- ·Skipping the dry-run because the spec 'looks right'. Simulated runs catch missing fields in seconds.
- ·Going live with zero gates on day one, then losing team trust to one bad write.
Go-live checklist
- [ ]Outcome named
- [ ]Webhook verified
- [ ]Spec dry-run clean
- [ ]At least one gate
- [ ]Owner assigned in Slack
Takeaway
A room is worth shipping the moment one step is autonomous — you do not need the whole loop on day one.
Read next
Signal from the field
What operators say
"We shut down three internal 'AI workflow' projects the week we adopted this. The room is the roadmap."
Priya Rao
Head of Ops, Fintech Series B
"Our KYB queue used to be 42 hours. It's now 90 seconds and my analysts do actual analysis."
Marcus Bell
Compliance lead, neobank
"The canvas made the workflow legible to legal and finance. That was the unlock."
Sofia Martins
COO, HR-tech scale-up
Changelog
Shipping every week
- Nov 12canvasInfinite canvas v2 — pan/zoom & mini-map
- Nov 08infra8-key Groq rotation with per-key cooldown
- Nov 04memoryMemory: outcome-scored playbooks (beta)
- Oct 29gatesSlack approval gates + audit export
- Oct 22runsCost caps per run + circuit breaker
Pricing
Priced per outcome, not per seat
Solo
$0forever
Room, canvas, console. 1 seat, 500 runs / mo, community connectors.
Team
Most rooms$490/ mo
5 seats, 25k runs, all connectors, approval gates, Slack + SSO. Cost caps.
Enterprise
Custom
VPC deploy, private key pools, SOC2, dedicated agent tuning, on-call SRE.
FAQ
Answers before you ask
+How is this different from a chatbot?
A chatbot answers; the room acts. Signals wake it up, plans get approved, runs execute in sandboxed tools, memory learns from outcomes.
+Which models power the agents?
Groq's llama-3.3-70b by default with 8 rotating keys and per-key cooldown. Bring your own OpenAI, Anthropic, or Gemini keys per workspace.
+Where does my data live?
Your workspace runs in an isolated tenant. Enterprise plans deploy inside your VPC. All tool calls are audited and replayable.
+How do I trust an autonomous agent?
Every plan is inspectable, every run is diffable, and every high-stakes action can be gated on human approval via Slack or email.
+What's the fastest way to see value?
Fork a playbook, connect one source, watch a run. Most teams have a first agent live before their trial coffee gets cold.
The evolution
From prompts to production fleets
01
Single agent
One prompt, one task, one answer.
02
Multi-agent team
Agents that plan, delegate, and check each other.
03
Autonomous system
A crew that perceives, decides, and acts across your stack.
4.1M
runs executed
128k
decisions in memory
99.2%
runs without human touch
340ms
p50 first token
120+
connectors live
8
rotating inference keys
Numbers refresh nightly from the production ledger.
status.replitagents.comTake the room live
Stop shipping features. Start shipping agents.
Fork a playbook, wire a signal, watch your first agent close a loop — in the next 5 minutes.