The runtime for AI agents

Build the agent.
Flux runs it.

An agent isn't one model call. It's a long-running system, and keeping one alive means building your own scheduling, routing, and recovery. Flux is the runtime that handles all of it, so you build the agent and it runs reliably, at a fraction of the cost.

−64%
inference spend
0
lost tasks
1 line
to integrate
plan retrieve ingest tool-call synthesize verify deliver write-report
FLUX · EXECUTION LOGLIVE
The missing layer

Every team is rebuilding the same plumbing.

Retries, scheduling, state, routing, recovery: the operational layer that keeps an agent alive in production. Today it's hand-rolled per team, per agent, and it breaks in the same handful of ways.

context_overflow

The window fills mid-task and the model silently drops the thing it needed most. The failure surfaces three steps later, somewhere else.

tool_error

A flaky API returns one 500 and the whole run is dead, even though 36 steps of perfectly good work already succeeded.

model_stall

The model loops, refuses, or times out at 2 a.m. Nobody notices until morning. The logs won't tell you which step, or why.

without a runtime · fail @ step 37 → replay all 40 steps → 2× cost, 0 answers
on flux · fail @ step 37 → resume @ step 37 → done
The runtime

Everything an agent needs to run, built in.

Scheduling, routing, recovery, memory: the operational layer you'd otherwise wire up by hand, provided as runtime primitives.

flux runtime8 PRIMITIVES
01Execution

Runs each unit of work to completion across models, tools, and steps.

02Scheduling

Orders and parallelizes steps, and knows what can run now and what has to wait.

03Routing

Sizes each step by difficulty on the fly and sends it to the cheapest model that can actually do it.

04Checkpointing

Captures state at every step, so work is never lost.

05Recovery

Resumes from the exact point of failure instead of restarting the task.

06Memory

Persists and compresses context across steps and across runs.

07Observability

Every step, cost, and failure named and visible live.

08Governance

Budgets, permissions, and limits enforced by the runtime, not your app code.

The runtime, running

Every step, named. Every failure, recovered.

A real run inside the Flux dashboard. Every step color-coded by outcome, every failure classified, every recovery costed.

flux dashboard · runsLIVE

task: market-research · run #147

ingestplan tool-callsynthesize write-report ✗deliver ✓
9/9
steps complete
1
failure recovered
$0.184
task cost
$0.42
saved vs replay

failure reasons · last 24h

reasoncountrecovered
context_overflow77/7
tool_error44/4
stall22/2
refusal11/1
budget11/1

Demo workload. Reproduce it yourself: pip install fluxcompute

The API

Recovery is one call.

agent.py
with client.task("market-research") as t:
    result = await client.messages.create(...)
    with client.step("fetch-pricing-data") as s:
        s.set_output(pricing_data)
    # a step fails deep into the task…

# instead of restarting from scratch:
response = await client.resume(t.task_id)
# → minimal context from succeeded steps, failed step re-routed

You write the agent's logic. Flux runs it: scheduling, routing, checkpointing, and recovery happen underneath, with no framework to adopt and no workflow to rewrite.

Cost

Reliability that pays for itself.

Flux sizes every step by difficulty on the fly and runs it on the cheapest model that can actually handle it, cutting inference spend without a line of model-selection code from you. The same judgment picks the right model on a retry.

direct-to-frontier
1.00×
with flux
0.36×

Demo workload, mixed difficulty · reproducible with pip install fluxcompute and your own keys

Team

Built by the people who run it.

I

Ishan Patwardhan

co-founder

Inference research at MIT CSAIL (hierarchical MoE, DAG-based token routing), SWE at Google on Gemini evaluation and context engineering, HPC systems at HPE. Researcher at Cornell Tech in agentic systems and hardware-software co-design.

ip259@cornell.edu
N

Niki Karanikola

co-founder

Two years shipping production LLM inference, dense retrieval, and RAG pipelines at Veolia, where she re-architected enterprise knowledge access from weeks to minutes, 6× faster reporting cadence, +20% NDCG.

nk699@cornell.edu

Our team's background: Cornell Tech · MIT CSAIL · Google · Veolia · HPE

We're onboarding early design partners.

Running agents in production, or trying to? We'll wire Flux under your stack with you, and you get direct access to the people who built it.

Become a design partner

Put the runtime under your agents.

Twenty minutes with the founders. Bring your worst-failing agent and we'll run it on Flux.

Book a demo