v0.4 · invite-only beta

Optimize any agentic workflow
against cost, speed, and quality.

Drop in a high-level task or an existing graph. AutoAW co-evolves topology, prompts, models, and tools — searching the Pareto frontier of your utility function so you ship the cheapest version that still hits your bar.

74.1% on GAIA · vs 69.2% Sonnet 4.5 baseline6.7× cheaper · vs same-quality GPT-5 single-agent38min to SOTA · 94 candidates evaluated
exp-7c4e · customer-support-copilot · gen 18/40
running
taskplannercs-4.5executor·ag2-flashexecutor·bg2-flashjudgehaikuanswer
quality
0.732
+0.041 vs gen 1
cost / run
$4.62
−84% vs gen 1
p50 latency
2.4s
−72% vs gen 1
02 · diff

From a brittle prototype to a Pareto-optimal pipeline.

AutoAW doesn't just tune prompts — it rewrites the graph. Below is the actual diff from the customer-support copilot experiment.

before · v0 · hand-written
taskreact-agentgpt-5docscrm.apiweb.searchcalcknowledgeanswer
quality
0.41
cost/run
$38.50
p50
12.4s
nodes
6 (1 model)
22 generations
after · gen 22 · AutoAW-optimized
taskplannercs-4.5exec·ag2-flashexec·bg2-flashjudgehaiku
quality
0.741
cost/run
$4.20
p50
2.4s
nodes
4 (3 models)
03 · ecosystem

The missing third layer.

DSPy compiles prompts. Inspect evaluates agents. Neither searches over architectures — AutoAW does.

DSPyprompt compile

Tunes instructions and few-shot examples for a fixed agent topology.

Inspecteval harness

Runs agents in sandboxed environments and scores them against benchmarks.

04 · architectures

Two paradigms. One optimizer. No assumptions.

Multi-agent committees and agentic glue are different bets on where reasoning should live. AutoAW makes neither bet upfront — it searches both and lets your fitness function settle the argument.

multi-agent committee
taskanswerplannergpt-5researchercs-4.5writercs-4.5validatorhaikucross-talkfeedback loop
LLM-to-LLM hand-offsfragmented contexthigh coordination overhead
wins when
·Security or permission isolation is required between agents
·Domain knowledge is too large for a single context window
·Task genuinely needs competing hypotheses (debate topology)
agentic glue + skills
taskanswerconductorclaude-sonnetweb_searchsql_querycode_execdeterministic tools — no LLM calls
one conductordeterministic toolsunified context
wins when
·Standard engineering, data extraction, or enterprise automation tasks
·Cost and latency are in the objective — the fitness function will find it
·One capable model can hold the full reasoning context end-to-end
DimensionMulti-Agent CommitteeAgentic Glue + Skills
LatencyHigh — each hand-off is a fresh LLM callLow — one LLM, tools run as fast code
ContextFragmented — state re-interpreted at each hopUnified — one context window, no drift
DebuggingComplex — trace across N reasoning chainsStandard — did the LLM call the skill? Did it run?
CostProportional to agent count × task lengthProportional to conductor reasoning only
05 · how it works

Five steps. One Pareto frontier.

STEP 01

Describe the task

Either give a high-level task — "triage support tickets and propose a draft reply" — or import an existing workflow (LangGraph, DSPy, your own Python).

# task.yaml
goal: "Triage support tickets,
  propose draft reply with citations 
STEP 02

Pick datasets & weights

Choose an eval dataset (yours or one of 18 included). Set the utility weights — quality, $/run, p50 latency. AutoAW does the rest.

qualitycostspeed
STEP 03

AutoAW searches

A search loop swaps models, splits/merges agents, edits prompts, prunes tools, and re-evaluates every candidate. The Pareto frontier fills in as it runs.

STEP 04

Export to AWP / AWF

Pick any point on the frontier and export the winning agent as a single portable graph — Agentic Workflow Protocol (AWP) or an AWF file, or plain JSON. Fork to keep optimizing.

$ autoaw export candidate-B --awf
✓ wrote candidate-B.awf
formats: awp · awf · json
STEP 05
Coming soon

Promote & deploy

Promote a candidate straight to a hosted, versioned endpoint — no infra to manage.

$ autoaw promote candidate-B
✓ deployed → exp-7c4e/prod
endpoint: https://run.autoaw.io/v1/...
06 · results

What it found in our last run.

Task: customer-support copilot, 1,400-ticket eval. Objective = 0.7·quality + 0.25·cost⁻¹ + 0.05·speed.

GAIA
74.1%
+5.3 vs. SOTA
baseline Sonnet 4.5: 69.2
Cost / run
$4.20
−6.7× cheaper
baseline GPT-5: $26.80
Latency p50
2.4s
−3.6× faster
baseline GPT-5: 8.7s
Generations to SOTA
22
≈ 38 min wall-clock
94 candidates explored
07 · case studies

Recent case studies.

Controlled benchmarks — hand-built workflows measured head-to-head against a strong monolith, not evolutionary runs. Code and per-task data released.

BigCodeBench · code generation

A workflow that runs its own tests beats a one-shot monolith.

On 150 hard BigCodeBench tasks we benchmarked a one-shot monolith against a workflow that executes each task's unit tests and repairs against the failures — then scored both on held-out tests the agent never sees. Base model gpt-oss-120b.

Pass rate
0.807
+13.4 pp vs monolith
one-shot monolith: 0.673
Win / loss
21 / 1
paired over 150 tasks
recovers 21 of 49 misses
Replicated
2 models
+6.0 pp on Qwen3-235B
held on a 2nd base model
Significance
10⁻⁵
McNemar p-value
not a small-sample fluke
09 · voices

What early users are saying.

Named results from teams putting AutoAW into their eval loop. Want to be here? Tell us your story.

We worked together in the past on AutoML, and now Andrey is following an interesting direction — how AutoML can be applied to agentic workflows. Very relevant and exciting!

Pavel Kordík
Pavel Kordík
PhD supervisor · AutoML & meta-learning researcher, Associate Professor, CTU Prague · Co-founder, Recombee
10 · pricing

Licensed for commercial use.

AutoAW is commercial software, licensed per team. Pricing is custom and based on your use case — workloads, integrations, and deployment model.

Not sure if your use case is commercial? Email us at hello@autoaw.app.
Request a demo & quotation

See AutoAW in action on your own workflow.

We'll walk you through a live demo and put together a custom quote for your team.
Request a demo