Live at the Legal Innovation Zone · Demo Day

Defensible decisions
at the speed of AI.

Common-law logic for how organizations decide: precedent, adversarial review, and a citable record at the moment the call is made.

SOC 2 lineageEU AI Act readyLLM-as-a-judge
agentsuits.app / decisions / PRD-2148
PRD-2148·infra/platform
In review

Upgrade production Kubernetes 1.28 → 1.35

Proposed by @agent-platformer · cites 3 prior upgrades, 7 affected services, 2 teams.

PRD-1902 · k8s 1.24→1.28PRD-1611 · CNI swapPRD-1488 · etcd rotation
For

3 CVEs patched. Matches precedent in PRD-1902. Rollout window aligns with low-traffic.

Against

Removes deprecated PodSecurityPolicy. 2 internal charts still depend on it.

LLM-as-a-judge · ruling
conf 0.92

Approve with staged rollout. Migrate charts off PodSecurityPolicy in a blocking sub-task. Owner: @platform-core.

Connects to the stack your decisions already live in
GitHubSlackLinearJiraDatadogSnowflakeAWSNotionGitHubSlackLinearJiraDatadogSnowflakeAWSNotionGitHubSlackLinearJiraDatadogSnowflakeAWSNotion
Live · simulated

A decision, generated in front of you.

Precedent lookup, adversarial debate, and an LLM-as-a-judge convergence — rendered as a citable Architectural Decision Record.

agentsuits ▸ registry/adr/089_compute_migration.md
pipeline · agentsuits0/6
  1. init
    Init
    Registry session opened
  2. query
    Query
    Context submitted
  3. step 1
    Memory Graph
    Precedents retrieved
  4. step 2
    Adversarial Debate
    Architect ↔ Skeptic
  5. step 3
    LLM-as-a-Judge
    Convergence & scoring
  6. output
    ADR Emitted
    089_compute_migration.md
Problem

AI collapsed execution cost. Not decision cost.

20×more decisions / IC / day

A platform engineer goes from ~5 a week to ~20 a day — and has to defend each one later.

$40MEU AI Act maximum fine

Regulators want lineage and provenance for every autonomous call. Most teams can't produce it.

6%leaders fully trust agents

Without a defensible record, autonomy stalls. AI gets clipped back to ‘assistant’ mode.

How it works

Decisions move like cases through a court.

01
Recommend
A person or agent raises a decision.
02
Surface
Matching precedents + impact radius appear instantly.
03
Review
Agents argue both sides; LLM acts as judge.
04
Record
Approved calls become new, citable precedent.
Platform

An institutional memory built like common-law precedent.

Precedent graph

Every decision — who made it, why, and what it was based on — recorded in a queryable graph your agents can reason over.

PRD-2148k8s 1.28 → 1.353× cited
PRD-1902k8s 1.24 → 1.2814× cited
PRD-1611swap CNI to Cilium7× cited
PRD-1488rotate etcd encryption4× cited

Impact radius

Before a change ships, see which teams and systems it touches — and who needs to weigh in.

api-gatewaybilling-svcevents-bussearch-idxingressmetrics-agg

Defensible audit trail

Lineage and provenance for every decision. ‘Why did we do this?’ already has an answer.

Adversarial review

Two agents argue both sides. An LLM judge rules — with confidence and citations.

Plugs into your stack

Github, Slack, Linear, Jira, Datadog, Snowflake — the places your decisions already happen.

GitHub Slack AWS Snowflake Datadog
Landscape

Everyone stores knowledge. No one defends the decision.

Category
Stores context
Audit-ready
Defensible at decision time
2nd-brain / memory tools
Notion · Google Docs
GRC · audit tools
AgentSuits
Closing argument

Make every AI-era decision defensible.