The agent framework where nothing is “done” until a grader returns a verdict — checked by real senses including eyes — and only verified work compounds into the fleet’s shared memory.
Every agent action is a hypothesis. One verdict bus unifies vision + tests + lint + types into a single pass / warn / fail — with attestation, so a hollow check can’t mint green.
A brain, eyes, a verdict bus, a fleet, and agents that build their own tools — each gated by the bus.
Memory that compounds — trust + provenance, consolidation, a held-out attested promotion gate, and lifecycle controls (pin · volatile-until-confirmed · TTL · correction chains). Only verified work graduates.
verel.memoryAgentVision as a perception organ (DOM / contrast / OCR grounded), feeding the verdict bus and the brain.
verel.sensesOne schema for every sense — advisory ceiling, grader attestation, scrubbed fingerprints, stuck/progress.
verel.verdictAgents managing agents — an LLM manager fans out; workers run in isolated git worktrees under budget.
verel.fleetDetect → scaffold → test → register → reuse, sandboxed with bwrap, admitted only on a passing attested eval.
verel.toolsmithSelf-healing pipeline with a deterministic rollback engine that never acts on advisory evidence.
verel.ciSituations a team actually hits — each a runnable script in examples/. The output
below is real, not mocked: an agent never decides “done”, a grader does.
Real pytest fails → an agent patches the source (never the tests) → the stage re-gates until the graders go green.
round 1: verdict=fail patched=['mathx.py','strx.py'] round 2: verdict=pass patched=[] healed=True terminated_on=passed$ python examples/demo_selfheal.py
Canary grader fails → deterministic git revert to the last good HEAD — and it refuses to act on advisory-only evidence.
canary verdict=fail rolled_back=True reverted c2852b1 → new HEAD 59ce62e advisory-only failure → executed=False (refused)$ python examples/demo_canary_rollback.py
Concurrent managers fenced by leases (a stale leader’s writes refused); a multi-repo change commits as an atomic saga — nothing half-applied.
8 tasks, 2 managers — ran once each: True stale leader A fenced off (token 1 < 2) saga: committed=[] compensated=['commit:api'] repos left landed: [] → atomic$ python examples/demo_distributed_fleet.py
Python + JS + Go tests, lint, types, a perf budget, a security scan, and IaC / cloud-IAM — catch dangerous Terraform / Kubernetes-RBAC grants before apply across AWS/GCP/Azure — all on one verdict schema, one gate.
Go inner-loop: test [-] TestLogin failed
JS pre-merge: test [-] submit posts the form
Python: security [-] B602 shell=True
perf [-] p95_ms 240 > budget 150
$ python examples/demo_polyglot_ci.py
detect → scaffold → test → register on a passing eval, then jailed to the syscalls it earned — a socket it never exercised is refused at the kernel.
learned 28 syscalls → enforced 71 tool runs under its jail: 5 socket() under the capability jail: REFUSED — Operation not permitted$ python examples/demo_capability_jail.py
Recall down a self→team→org→global lattice; a peer’s claim re-verifies before it’s trusted; leader-fenced HA with quorum reads that survive the leader being down.
agent-B's belief (my check fails): stayed CANDIDATE — trust did not travel quorum read, leader DOWN → still returns 'restart the worker pool' (freshest)$ python examples/demo_shared_brain.py
Extract facts from a conversation like Mem0/Engram — but a fact only compounds after it’s graded: it stays CANDIDATE until attested or corroborated by ≥2 authenticated principals, so a one-off (or an attacker repeating a lie) is never trusted. Recall is BM25, token-budgeted, graded-first, and fenced as untrusted data — so you stop replaying the whole brain into every prompt (~80% fewer tokens/turn in the demo).
single say-so → trust=CANDIDATE (not trusted) 2 authenticated principals → trust=VERIFIED attacker repeating it N× → stays CANDIDATE recall budget=12 → used=8, dropped=2 (fenced)$ python examples/demo_memory.py
AgentVision is the eyes; Verel is the brain. The eyes perceive a rendered artifact and grade it — including does it match what we set out to build? — then hand a clean signal up the optic nerve. The brain decides with grader attestation, and only verified work compounds into memory. The eyes can also watch over time — temporal verification of playback / loading / liveness for streaming UIs, live dashboards, and now local video files & animated GIFs graded frame-by-frame. A deterministic video stall or a dead/static export gates the bus to FAIL, and verified playback compounds into memory. They ship and version independently (pip install agentvision, pip install verel), but in sync.
The five organs and the eval-driven loop — hand-drawn, and verified clean by the eyes Verel ships.
Rendered and verified by the eyes Verel ships — Verel gates its own marketing.
Real PyPI downloads by country — last 30 days, via BigQuery. Verel ~15.8k · AgentVision ~6.1k downloads across 60+ countries.
Top countries by 30-day PyPI downloads · snapshot 2026-06-25
CI gate in one step (uses: amitpatole/verel@v1.9.4) — tests + lint + types in one verdict — a pre-commit hook, or verel-ci check in any script. For agents: verel-mcp exposes the verdict bus + memory, with AgentVision as the eyes.