T
E
TENET
E
T
The Agent Operating System

Memory that compounds.
Agents that learn.
Ships while you sleep.

Every other AI tool forgets when the session ends. TENET gives agents persistent memory, an eval-gated learning loop, and multi-agent coordination — so each session builds on the last.

$npm install -g @10et/cli·tenet init
30day
release cadence
95→14
processes cleaned
15GB→1.3GB
swap reclaimed
scroll to explore
01 — Memory

Context that
compounds

Every session writes structured journal entries — features, decisions, fixes, discoveries. Semantic search finds them instantly. The next session starts with full context, not a blank page.

0
memories indexed
0
sessions journaled
0
training tuples
→Every action writes a journal entry (feature, decision, fix, discovery)
→Semantic search with BM25+ hybrid — finds decisions from months ago
→Graph edges: updates, contradicts, related_to, caused_by
→New session starts with P1 issues + recent journal + project context injected automatically
memory.db — live journal349 entries
decision

Auth flow: offline-first, opt-in cloud

2h ago
0.94
feature

Recipe system — tenet recipe run/list/show

3h ago
0.91
fix

HOME .git guard — prevents macOS permission flood

4h ago
0.88
discovery

Goose recipe format: self-contained YAML tasks

5h ago
0.85
feature

Fleet recipes: visa-red-team, full-gauntlet

6h ago
0.92
decision

Trust ladder L1→L4 — graduated CI automation

8h ago
0.87
fix

Zombie process cleanup — 95→14 node processes

10h ago
0.96
milestone

Beta gate deployed to 10et.ai — password protected

12h ago
0.83
$tenet memory search "auth decision"▮
02 — Learning Loop
tenet eval status — live scorecard
code-quality
0.426 / 0.80
12/216 kept (5.6%)✓ gradient
test-coverage
0.120 / 0.30
8/180 kept (4.4%)✓ gradient
cli-speed
3,635ms / 2,000ms
4/90 kept (4.4%)✓ gradient
memory-recall
0.000 / 0.80
0/0 kept (—%)
tenet peter agent code-quality --rounds 10▮

Agents that
get better

Each agent has a metric, an eval script, and a learning loop. It tries changes, measures them, keeps what improves, reverts what doesn't. Runs overnight. You wake up to improvements.

01Write specDefine what good looks like
02Generate evaltenet build --spec your-spec.md
03Agent iteratesTry → eval → keep/revert
04PR createdOnly if metric improves
05Policy updatesRL learns what worked
↺ repeat nightly

Granularity of feedback determines speed of convergence. Monolithic eval = 7% keep rate. Decomposed binary checks = 100% convergence in 1-3 rounds.

03 — Self-Driving Pipeline

Ships while
you sleep

The trust ladder is how you go from "AI reviews PRs" to "AI ships daily releases." Each level adds more automation. You decide when to climb.

1
AI Reviewactive

3 models review every PR. Read-only. Comments findings.

$ tenet ci setup --sentinel
Anthropic + OpenAI + OpenRouter reach agreement before flagging
2
Eval Scoringactive

Agent evals score every PR. Passes threshold or requests changes.

$ tenet ci setup --eval
eval-on-pr.yml runs on agent branches — keeps if metric improves
3
Auto-Merge

Merge when eval passes + review agrees. Human can still override.

$ enable branch protection
4
Auto-Publish

npm publish on every merge to main. Every day is a release.

$ tenet ci enable release
30-day goal

Ship a new @visa/cli release every day for 30 days. After 30 days, open-source TENET with Visa CLI as the #1 case study.

Visa-Crypto-Labs/Visa-CLI
#105

cleanup: remove made-with-visa prototype

Apr 5sentinel ✓
#103

feat: auto-open Studio dashboard during generations

Apr 5sentinel ✓
#102

fix: reject conflicting pxlwall_card modes

Apr 3sentinel ✓
#101

feat: update pxlwall_card for new API + video gen

Apr 3sentinel ✓
#100

fix: resolve 4 code scanning alerts

Apr 3sentinel ✓
CI Score38%
secrets missing100% = daily auto-ship
04 — Multiplayer AI

Every feature runs
the gauntlet

Positioned agent fleets attack every Visa CLI feature from different angles. The CRO asks about kill switches. The GC asks about cardholder agreements. The fraud attacker finds the velocity exploit.

$ tenet recipe run full-gauntlet.yaml --param feature="AI agent payments"
Panel sources
visa-skeptics(5 agents)
competing-builders(5 agents)
adversarial-consumers(5 agents)
regulators(5 agents)
🛡️
Patricia Chen
Chief Risk Officer
CONDITIONAL
Concern raised

“What's the 99th percentile loss scenario? I need a kill switch — halt all AI agent transactions globally in under 60 seconds.”

Requirement to proceed

Quantified risk model + emergency halt API

Auto-filed GitHub Issue
P1securityneeds-context

[Red Team — Patricia Chen] Quantified risk model + emergency halt API

→Findings auto-filed · overnight agents address blockers · next release re-runs gauntlet
05 — Recipes

One command.
Any task.

Recipes are self-contained YAML task definitions — parameters, eval criteria, and a prompt in one file. Share it, run it, import from Goose, build your own.

Compatible with Goose (35K ⭐)
$ tenet recipe import analyze-pr --source goose

Any of Goose's 48 community recipes imports instantly — and gains eval criteria + learning that Goose doesn't have.

terminal
$ tenet recipe run daily-release-check.yaml --param repo=Visa-Crypto-Labs/Visa-CLI
  Daily Release Check
  [beta] Visa-Crypto-Labs/Visa-CLI

  Checking CI status...
  ✓ ci.yml — success
  ✓ tenet-sentinel.yml — success  
  ✓ eval-on-pr.yml — success

  Checking P1 blockers...
  ✓ 0 open P1 issues

  Recent merges (changelog):
  → #105 cleanup: remove made-with-visa prototype
  → #103 feat: auto-open Studio dashboard
  → #102 fix: reject conflicting pxlwall_card modes

  ────────────────────────────
  ✅ READY TO RELEASE
  Suggested bump: patch → v1.0.3
  ────────────────────────────
tenet_journal_write called✓
# daily-release-check.yaml (excerpt)
version: 1.0.0
maturity: beta
eval:
- check: gh CLI available
command: which gh
expect_exit: 0