Skip to main content
AaaS·TripleA · agent reliability infrastructure

Deterministic superpower
for any AI agent.

plan → execute → gate → audit pipeline DAGsupervisorRL retry · goal-check fail
01
plan
02
execute
03 · approved
gate
04
audit
~/audit — chain
[ctx] plan loaded · 5 nodes
[evt] node 01 plan ✓
[evt] node 02 execute ✓
node 03 gate approved
sha256:9f3a8b3c…
↑ 4c81de2e…
↑ b2de0791…
↑ e1f4a209…
chain verified

wrap Claude, GPT, open-weight, or your custom agent in a pipeline that can't drift, skip steps, or lie about what it did.

MIT core · SOC-2-ready license · self-host or cloud · works w/ any LLM

same engine · pick your tradeoff

One pipeline, two runners.

plan.yaml forks into interactive and programmatic runners
plan.yaml
▸ MODE 01
Interactive.
your AI chat is the worker · zero extra tokens · one terminal · low-budget · fast iteration
aaa run plan.yaml --mode interactive
▸ MODE 02
Programmatic.
headless subprocess workers via SDK · full isolation · production-grade · parallelizable · billable per task
aaa run plan.yaml --mode programmatic --workers 8
▸ MODE 01
Interactive.
your AI chat is the worker · zero extra tokens · fast iteration
aaa run --mode interactive
▸ MODE 02
Programmatic.
headless workers · isolated · parallel · billable per task
aaa run --mode programmatic --workers 8

the reliability gap

What changes when
an agent runs through AaaS·TripleA.

▸ WITHOUT AaaS·TripleAuntrusted
untrusted agent claims done while skipping workai"done."skipped testsforgot todo
▸ WITH AaaS·TripleAverified
gated pipeline: plan, execute, gate, audit, replayplnexcgteaudrep
01
AI says "done" — you check the diff.
per-node supervisor decision + structured TODOs.
02
AI forgot the tests.
pipeline blocks until test node passes.
03
What did the agent do at 3am? — guess.
HMAC-chained timeline. replayable.
04
Every session re-invents the wheel.
YAML plans = version-controlled library.
05
Locked to one model.
swap Claude / GPT / local — same gates.
06
n8n / Zapier bill per run, no AI discipline.
self-host engine. AI-native gates built in.

the protocol

Setup once. Run forever.

01
Out-of-scope setup.

declare your orchestration surface once — tools, MCPs, per-node LLM, agent assignments. lives outside the run; the run never re-decides it.

$ aaa init && vim agents.yaml
02
Deterministic execution.

/autonomous walks the DAG node-by-node. hard gates block. soft gates warn. RL retries with structured feedback until goal-check passes — or the node escalates.

$ aaa run plan.yaml --verbose
03
Audit & replay.

every node emits a signed event. the full run is an HMAC-chained timeline. replay byte-identical, diff against a prior run, or hand to compliance.

$ aaa replay run_a91c.log --diff prev

features

Six modules, one pipeline.

▸ 01
Deterministic execution
DAG + RL + goal-check. no free-form drift.
▸ 02
Visual editor
n8n-style canvas for nodes / gates / supervisors.
▸ 03
Plan generator
describe the outcome; AaaS·TripleA scaffolds the flow.
▸ 04
Hard & soft gates
block on test fail, warn on lint, escalate on decline.
▸ 05
Model-agnostic
Claude · GPT · Gemini · Llama · local 7B. per-node.
▸ 06
Replayable audit
HMAC-chained log. reproduce any run.

positioning
n8n, Make, and Zapier are workflow tools that bolted AI on top.
Cursor and Copilot are agents without a discipline layer.
AaaS·TripleA is the discipline layer — AI-native gates, supervisor decisions, signed audit — built for the assumption that the worker is an LLM and will lie about being done.

pricing

Three tiers. No "contact us".

PERSONAL
free.
self-host. MIT core. unlimited interactive runs.
TEAM
$X /seat/mo
programmatic mode. hosted supervisor. shared plan library.
COMMERCIAL
license
SOC-2-ready. audit export. on-prem.
▸ aaas.love
Stop trusting
"done". Start
verifying it.
aaas.love · github.com/aaas/aaa · docs · changelog · status
withlove· aaas.love
built by operators who got burned by agents at 3am.