Capability
is not authority.
Founder · AI infrastructure engineer · Lagos, Nigeria
I build governed systems for autonomous intelligence.
Agent frameworks, persistent runtimes, memory, financial controls, and evaluation infrastructure. Together, they form a connected system where capability never silently becomes authority.
Snorkel AIExpert contributor & reviewer
1st placeAgentic Commerce on Arc
PublishedPolicy-constrained execution paper
REFERENCE EXECUTION TRACE POLICY BOUNDARY
INTENTAgentrequest()
CONTROLPolicy kernelverify · constrain
EXECUTIONBounded actionevidence attached
AUTH ≠ AUTHORITYDIGEST CHECKDECISION RECORDEDSTEP 01 request.received ✓
STEP 02 policy.evaluate ✓
STEP 03 action.authorized bounded
AGENT RUNTIMES◆DURABLE MEMORY◆POLICY ENFORCEMENT◆LONG-HORIZON EVALUATION◆IMMUTABLE EVIDENCE◆HUMAN + AGENT WORK
01 Operating thesis
Useful autonomy needs a boundary.
The hard part is no longer getting an agent to act. It is deciding what that action means, whether it is allowed, what evidence must exist, and who remains accountable when the workflow lasts for hours.
My work focuses on that control plane: the infrastructure between an intelligent request and a consequential real-world outcome.
Authentication
is not authorization.
Submission
is not acceptance.
A payment request
is not permission to execute.
02 One connected body of work
Not a project list.
A systems stack.
Each component solves a different failure mode in autonomous systems. Together they form a path from reasoning to durable, governed execution.
THINKREMEMBERPERSISTGOVERNEVALUATE
INTERACTION LAYERJarvis ProtocolHuman ↔ agent work
CONTROL PLANEOmniClaw CoreBuyer policy + payment execution
SETTLEMENT SERVICEHosted FacilitatorSeller verification + settlement
EVIDENCE PLANEWorkstreamEvaluation + contribution
AGENCYOmniCoreAgent
RUNTIMEOmniDaemon
MEMORYOmniMemory
ORIGIN · JAN 2026OmniAgentPayThe hackathon-winning developer SDK demonstrated policy-controlled spending for autonomous agents.
evolved into→
CURRENT ARCHITECTUREOmniClawThe architecture separates buyer-side policy control from seller-side hosted settlement.
01 / Agent harnessOpen source
OmniCoreAgent
The open production agent harness for Python: an explicit runtime boundary around a model with parallel tool batches, structured observations, MCP and local tools, memory, workspaces, guardrails, subagents, background tasks, and REST/SSE serving.
02 / Cognitive memoryOpen source
OmniMemory
A production memory framework that uses dual-agent synthesis, self-evolving conflict resolution, composite retrieval scoring, and app/user/session isolation to turn interaction history into useful memory.
03 / RuntimeOpen source
OmniDaemon
A universal, framework-agnostic event-driven runtime that runs each agent in an isolated supervised process with auto-recovery, health monitoring, retries, dead-letter queues, metrics, and distributed coordination.
04 / Buyer controlActive build
OmniClaw Core
The successor to OmniAgentPay: buyer-side policy-controlled payment infrastructure with a financial policy engine, Python SDK, agent CLI, wallet and payment routing, x402 execution, Gateway readiness, idempotency, simulation, and payment-intent controls.
05 / Seller settlementPrivate alpha
OmniClaw Facilitator
A hosted x402 settlement service for sellers: it resolves seller policy from API keys, verifies payment payloads, settles supported rails, persists settlement state, and gives operators an auditable control plane.
06 / CollaborationProtocol Alpha v0.1.0
Jarvis Protocol
The protocol for governed collaboration and shared learning between HumanWorkers and AgentWorkers. It defines durable WorkSessions, reviewable requests, attributable contributions, governed learning, and portable evidence.
07 / Contribution truthv0.1 active development
Workstream
Governed contribution infrastructure that turns project-defined tasks, immutable submissions, deterministic checks, and authorized review into durable ContributionRecord facts for work performed by humans, agents, or both.
03 Selected experience
The work behind
the systems.
My current research direction did not begin with abstractions. It grew from designing protocol-driven infrastructure under real engineering constraints, then learning how frontier systems must be evaluated when the work lasts beyond one response.
TASK AUTHOR · REVIEWER · FRONTIER MODEL EVALUATION
Designing evaluations that reward the real work, not a convincing shortcut.
I create and review technically rigorous evaluations for advanced AI systems, with a focus on workflows that must remain correct across multiple steps, tools, and execution attempts.
- Create technically rigorous coding and agent-evaluation tasks for terminal-based and long-running execution workflows.
- Validate model behavior against detailed rubrics and specifications, including comparative analysis across repeated execution runs.
- Review tasks created by other contributors for benchmark quality, technical correctness, and alignment with project standards.
ELECTRIC-VEHICLE INFRASTRUCTURE
Building the backend control plane for connected charging stations.
Worked on the end-to-end backend and infrastructure design of OCPP charging-station management systems, balancing real-time protocol communication, concurrent workloads, and operational simplicity.
- Built Python services with Django REST Framework and FastAPI, using WebSockets for OCPP communication and RabbitMQ/Celery for event-driven background work.
- Improved database schemas and queries for transaction-heavy charging workflows.
- Moved internal deployments from Docker Compose to Docker Swarm, enabling rolling updates across project environments.
- Designed GitHub Actions pipelines for automated builds, tests, and containerized deployments.
REAL-TIME PROTOCOL SYSTEMS→DISTRIBUTED EXECUTION→GOVERNED AUTONOMY→LONG-HORIZON EVALUATION
04 Current build
WORKSTREAM · IN DEVELOPMENT
Preserving the truth behind every evaluation decision.
In an evaluation system, a file is not just a file. It may be the instruction that governed a task, the bytes a contributor submitted, or the evidence a reviewer accepted.
“If the evidence can be silently replaced, the integrity of the evaluation lifecycle collapses.”
REFERENCE LIFECYCLEDESIGN MODEL
CONTENT IDENTITYsha256:server-computedbyte count + immutable identity
✓SubmittedExact bytes recorded
✓CheckedDeterministic evidence
✓ReviewedDecision bound to version
04AcceptedContribution created
Same path ≠ same artifact
05 Proof, not posturing
01
GOOGLE GEMINI CHALLENGE · JAN 2026
First hackathon.
First place.
OmniAgentPay was the early implementation. It won the Google Gemini Challenge at the Agentic Commerce on Arc Hackathon. The work then evolved into OmniClaw: a buyer-side policy-controlled core and a separate hosted x402 facilitator for seller settlement.
1STPLACE
Policy-Constrained Financial Execution for Autonomous Agents
A formal architecture for separating an agent's ability to request payment from the authority to move money through policy, evidence, and bounded execution.
DOI10.5281/zenodo.20487323
Read the paper 06 Research direction
Evaluating agents that work beyond the demo.
Long-horizon agents fail in ways a single final answer cannot reveal. I am exploring benchmark infrastructure that evaluates the entire trajectory: environment, action, artifact, evidence, and acceptance.
Task validity
Can the task actually be solved from the supplied evidence, pinned environment, and stated contract?
Evaluator integrity
Do the checks verify the real work, or only a self-consistent output that an agent can fabricate?
Long-horizon evidence
Can every consequential decision be traced across attempts, tools, artifacts, reviews, and revisions?
Reproducible acceptance
Would the same exact bytes, policy version, and environment justify the result again later?
TRUSTWORTHY EVALUATION=PINNED ENVIRONMENT+REAL ORACLE+GAME-RESISTANT CHECKS+TRACEABLE EVIDENCE
AB/IALAGOS · NG
07 The engineer behind the systems
I care about what happens after the model says “done.”
I'm Abiola Adeshina, a software and AI infrastructure engineer based in Lagos. I build the operational infrastructure that helps autonomous AI systems persist, coordinate, recover from failure, and execute reliably beyond demos.
At Snorkel AI, I create and review technically rigorous evaluations for frontier agents, focusing on solvability, reproducibility, coverage, grader robustness, and resistance to shortcuts. The same systems discipline I developed in real-time EV infrastructure now shapes how I evaluate long-horizon agent behavior.
Across agent runtimes, orchestration, memory, evaluation, and policy-controlled execution, I follow the state, make failure observable, and keep consequential actions inside clear trust boundaries.
CORE DISCIPLINEDistributed systemsRuntimes, orchestration, reliability
FOCUSGoverned autonomyEvaluation, policy, financial controls
EARLIER SYSTEMSEV chargingOCPP charging-station management systems
BUILDER ROLEFounderomnirexflora-labs
Available for ambitious infrastructure work
Building agents is easy.
Making them accountable is the work.
If you are working on agent infrastructure, evaluation systems, governed execution, or long-horizon reliability, let's compare notes.
Email Abiola