Whose intelligence?

Building Agentic AI in Africa without frontier-model dependency.

--Days:--Hours:--Minutes:--Seconds

FormatForums & Dialogues

Length90 minutes

VenueDLI 2026

Hosted byML Collective (Africa)

Workshop outputOpen Preprint

Overview

Abstract

The promise of Agentic AI is seductive: autonomous systems that perceive, plan, and act, thereby reducing the friction between human intent and outcome. However, Agentic AI is arriving in Africa on someone else's terms. Building agentic systems today almost inevitably means building on frontier models trained predominantly on Western data, optimized for Western languages, and aligned to Western values; yet deployed to negotiate, decide, and act on behalf of African users. Every API call, every default behavior, every quietly embedded assumption carries a worldview that was not entirely built with Africa in mind.

This session takes a positional stance: the current default is insufficient and, in some cases, harmful. Through a workshop challenge, positional talks, and a structured fishbowl dialogue, we pressure-test four tensions — AI Safety & Policy, Infrastructure & Architecture dependency, Cultural Alignment, and Value Extraction & Capability Trap — working toward a community-driven position on sovereign agentic AI for Africa. The conversation will not end in the room; outputs will be compiled into an open preprint.

Goals & Objectives

This session aims to:

→Critically examine the assumptions embedded in building Agentic AI on top of frontier models in the African context.
→Move the African Agentic AI community beyond adoption conversations toward ownership and sovereignty conversations.
→Surface honest, grounded perspectives on the trade-offs between feasibility and dependency when building agentic systems under resource constraints.
→Contribute a community-driven position on what responsible, contextually-grounded Agentic AI development looks like for Africa.

Expected Outcomes for Participants

→A clearer framework for evaluating when and how to use frontier models in agentic AI tasks versus investing in local and compute-efficient capacity.
→Exposure to concrete examples of agentic AI suitable for African realities, grounding the theoretical tensions in practice.
→A structured post-session report detailing key positions, points of divergence, and resulting recommendations, published as an open preprint.

Topics of Interest

The Four Tensions

Today, building agentic AI almost inevitably means building on top of frontier models trained predominantly on Western data, optimized for Western languages, and aligned to Western values. When we deploy agents that negotiate and decide on behalf of African users, we are not just plugging into an API — we are embedding a particular worldview into the infrastructure of African lives.

This session takes a positional stance: the current default is insufficient and, in some cases, harmful. We are not arguing against using frontier models. We are arguing that we have not been honest enough about what we are trading away when we do.

Tension 01

The AI Safety & Policy Problem

Safety guardrails and governance frameworks for agentic AI are predominantly designed for Western contexts. Do they hold when applied to African languages, cultures, and realities? Are the policies we are adopting encompassing enough for us; to what degree are we included or not? How do we ensure that we reduce catastrophic risks arising from AI Agents while factoring in our cultural context?

Tension 02

The Infrastructure & Architecture Paradox

Africa faces real compute and infrastructure constraints, and the most feasible path to Agentic AI currently leads straight to dependency on the very systems that do not fully capture our nuances and context. The infrastructure required to train frontier-scale models from scratch is unavailable to most African institutions. Yet the alternative — permanent reliance on external APIs — carries its own risks. How do we pursue agentic capability in a constrained setting without necessarily deepening dependency on systems built elsewhere?

Tension 03

The Cultural Alignment Problem

Frontier models are not culturally neutral, and the gaps they carry across language, context, and epistemic tradition do not disappear when you fine-tune on a small Yoruba or Amharic dataset. The cultural assumptions baked into these systems — about what constitutes a valid question, a good answer, a trustworthy source — reflect the communities that produced the training data and shaped the RLHF feedback. Whose intelligence are we building on?

Tension 04

The Value Extraction & Capability Trap

When African developers and users interact with frontier models, they generate data and feedback that flows back to improve those models — for the benefit of labs elsewhere. We are, in effect, subsidizing the refinement of systems that were not built for us. The more African users adopt these systems, the more valuable they become to their creators, while the structural gap widens. Every time we default to a frontier model API instead of investing in local capacity, we make it harder to ever not need that API. The shortcut today becomes the ceiling tomorrow.

This session does not pretend to resolve these tensions. It convenes researchers, practitioners, and policymakers to pressure-test them; to argue openly, rigorously, and without deference to the hype. The question is not whether Africa should build Agentic AI. The question is whether we are being clear-eyed about what we are building, for whom, and on whose terms.

Activities

Session Format & Audience

Target Audience: This session is designed for practitioners, startups, researchers, and policymakers at intermediate to advanced career stages who are building, studying, or shaping agentic AI systems in African contexts. Participants are expected to have familiarity with LLMs and agentic system concepts; this is not an introductory session. Startup founders navigating real infrastructure constraints will find particular value in the workshop challenge solutions and fishbowl segments.

Opening TeaserA live provocation designed to immediately surface the tensions and set the tone for honest, critical dialogue.
Workshop ChallengePresentation of hackathon solution, detailing an example of agentic AI built for African realities under resource constraint, through a workshop challenge/hackathon.
Positional Q&AShort, provocative framings of each tension, designed to create productive disagreement rather than consensus.
Fishbowl DialogueThe core engagement mechanism; participants are not an audience but active contributors to a structured exchange.
Live Position CaptureKey positions and disagreements are documented throughout, feeding directly into the post-session preprint.

Workshop Challenge

Challenge Details

Coming Soon

Full challenge brief, submission guidelines, and prizes will be posted here.

References