Verel — Verified Agents 👁️🧠

Problem: AI agents declare work "done" on their own say-so — shipping broken UIs, failing tests and unverified claims they can't actually check. Result: Verel makes "done" a verdict, not an opinion — every action is graded by real senses (including eyes, via AgentVision), and only verified work compounds into the fleet's shared memory.

pip install verel
verel doctor                 # check your environment
verel heal --repo .          # self-healing CI: failing tests → agent fixes → green

One verdict bus unifies vision + tests + lint + types into a single pass / warn / fail, so progress, "done", and what compounds are all decided in one place — with grader attestation (a signed receipt of what actually ran) so a hollow check — a grader that claims success with no evidence — can't mint green.

Is Verel for you?

• Agents that write code, and you need a grader that re-runs the real tests so the agent can't fake "done"? → yes. Start: 5-minute tutorial.
• A fleet of agents sharing one brain, where a hallucination (or one bad actor) must not become trusted memory? → yes. Start: Memory in 5 minutes.
• Agents that render UIs but never look at them (overflow, contrast, 404s, video stalls)? → yes, with AgentVision as the eyes.
• Paying too much for context because you replay the whole memory into every prompt? → yes — budgeted, graded-first recall cuts a 40-fact brain ~80% (679→135 tok/turn). Cost →
• Gating agent-authored 5G RAN/Core changes (a slice missing from NSSF, a gNB TAC no AMF serves, a KPI regression)? → yes — deterministic, offline, one rule body grades Helm and NETCONF. Telecom RAN / 5G Core →
• Just want a memory layer for a single agent with a human curator? → Verel works, but Mem0 is simpler. Here's the honest when-to-use comparison.

The six organs

Organ	Module	What it does
🧠 Brain	verel.memory	Memory that compounds — only verified facts/skills graduate (held-out, attested promotion gate); lifecycle controls keep it from becoming a junk drawer. Pluggable backend (VEREL_MEMORY_BACKEND): local SQLite, a shared hosted brain, or an external DB.
👁️ Eyes	verel.senses	AgentVision as a perception organ (DOM/contrast/OCR grounded, intent conformance, temporal watch) feeding the verdict bus and the brain.
⚖️ Verdict bus	verel.verdict	One schema for every sense — advisory ceiling clamp, grader attestation, strict-subset stuck/progress.
🚁 Fleet	verel.fleet	Agents managing agents — LLM manager fans out, workers in isolated git worktrees, each gated by the bus.
🔧 Tool-smith	verel.toolsmith	Agents build their own tools, sandboxed, admitted only on a passing attested eval.
♻️ Agent-run CI/CD	verel.ci	Self-healing pipeline + deterministic rollback that never acts on advisory evidence; graders span Python/JS-TS/Go, perf, security, mutation, spec/intent — and IaC / cloud-IAM / Kubernetes-RBAC (catch dangerous Terraform/cloud grants before apply).

Eyes & brain

AgentVision is the eyes; Verel is the brain. The eyes perceive and grade (including does it match what we set out to build? and does the video actually play?); the brain decides with attestation and compounds only verified work into memory; then the eyes look again.

Next

• Memory in 5 minutes — extract → grade → recall, offline, no key. The fastest standalone win if you came for the memory.
• Verel vs Mem0 / Engram / Zep — honest when-to-use, and a "coming from Mem0" mapping.
• Try it yourself — a from-scratch, copy-paste walkthrough (no API key): catch a real bug, fix it, watch Verel remember it so it can't come back.
• Start here — zero to your first verdict in two minutes; the one core idea.
• Get started — install, the gate, CI/agents adoption.
• Install & extras — the light base wheel + every optional extra, and when to use it.
• Trust model — what's cryptographically guaranteed (signed receipts, isolation) and what isn't.
• 5-minute tutorial — gate a repo, heal failing tests, watch a bug get remembered.
• Use cases — where Verel fits: agent loops, CI/CD, fleets, shared memory.
• Real-world scenarios — runnable demos with real captured output for each.
• Architecture — the six organs and the eval-driven loop.

Install: pip install verel · Source: GitHub · Package: PyPI · License: MIT.