Agriculture: The Extension Backlog — Agentic AI for Africa Innovation Challenge

← Agentic AI for Africa Innovation Challenge

Overview

Call it a county office, a district office, an LGA desk, or a municipal extension service — whatever your local government calls it where you're building from, the same two or three people are responsible for turning weather, pest, and price signals into advice for thousands of smallholders they can't all visit this season. One extension officer described her week like this: rainfall estimates shift, a pest alert lands in a WhatsApp group at 6am, and the price at the nearest aggregation centre moves enough to change what a farmer three villages over should plant, spray, or sell — all before she's finished her first round of visits.

This track isn't written for one country's administrative map. Build for whatever the local agricultural office looks like where you are — a two-person desk in a small municipality is as valid a target user as a better-resourced district office. What matters is the workflow: signals in, household-level advice out, an officer in between.

1 season

Cross-checking rainfall, pest reports and prices against each household's plot, by hand

~15 min

What your agent should take to do the same work, per household cluster

Officer decides

Your agent stops here and hands over a sourced advisory, not an instruction

One rule that is not negotiable: your agent drafts the advisory, it never sends it to a farmer and never places an input order or a loan request on their behalf. A named officer at the local agricultural office approves before anything reaches a household — by SMS, USSD, or a home visit. Every recommendation must cite the rainfall, price, or plot record it came from; an unsourced planting recommendation is worse than no recommendation at all.

Choose one theme

Climate-smart advisory
Rainfall and pest signals against a household's plot history and crop calendar. When to plant, when to spray, when to hold.
Market access & price intelligence
Price outliers across aggregation centres, matching a harvest to the buyer paying the most within reach.
Input & credit risk
Matching a household's yield history and repayment record to an input package or a credit offer it can actually service.
Post-harvest & traceability
Tracking a harvest from farm gate to aggregator to buyer, and flagging where loss is happening in the chain.

Jargon, defined up front

MCP — Model Context Protocol: An open standard for exposing tools to an AI model. You write a small server; any agent can call its tools.
Agentic: Plans a multi-step task, calls tools, checks itself, recovers from failure. Not a chatbot.
Agroclimatic zone: A geography grouped by rainfall and soil so crop advice generalises within it, not across it.
USSD: The menu-driven mobile interface (dial *xxx#) most extension advisories and mobile money still run on in rural Africa.
Open-weights: A model you can download and run yourself. Qwen, Llama, Gemma, Mistral, Aya.
Human in the loop: The agent pauses, and a named person approves before anything consequential.

Requirements

Projects missing any item in either list are ineligible for prizes.

What to build

Your own MCP server. Three or more different tools, and at least one that does something — drafts an advisory, files a flag, schedules a follow-up. Read-only is not an agent's hands.
One MCP server you did not write. Community, official, or vendor. One line on why it beats writing it yourself.
Open-source orchestration. LangGraph, CrewAI, Pydantic AI, smolagents, Letta, Goose, n8n or equivalent. No closed no-code builders.
One full task on an open-weights model. Household and yield records often cannot leave the country. Show a frontier model alongside if you like.
A logged tool call for every action, and a gate on every irreversible one — before a message goes to a farmer, before an input order or credit flag is raised. Inputs, outputs, timestamps, and a named human approving.
New work, built during the submission period. Existing open-source libraries are fine; an existing project resubmitted is not.

What to submit

A public code repository with an OSI-approved licence and a README that gets a stranger running in one command.
A demo video under 3 minutes, uploaded to YouTube or Vimeo and publicly visible. An unedited agent run with tool calls on screen, not slides.
A text description of what it does, which sub-theme it fits, and which office and workflow it serves. Around 300 words.
ARCHITECTURE.md — one page. Agent shape, MCP servers built versus borrowed, and why.
EVALS.md — eight or more test tasks with pass and fail results, plus one failure you did not fix and what you would try next.

Three tools that work beat nine that half-work. In five days, scope is the skill we are watching for.

Prizes And Awards

Mentorship and build resources
Top 40 — selected from all submissions
Every shortlisted team gets mentor support and the other build resources described in the challenge brief.
All expenses paid
Top 10 — selected from the top 40
The top 10 teams are fully funded to attend the MCP Conference in Nairobi this November.
Amount TBC
1^st
Amount TBC
2^nd
Amount TBC
3^rd
Amount TBC
Best Women-led solution

How selection works: the top 40 submissions are shortlisted from all entries and receive mentorship and build resources; from those, the top 10 are selected and all expenses are paid for them to attend the MCP Conference in Nairobi this November. One prize per team. Judges may decline to award a prize if no submission meets the bar.

Judging criteria

Criterion	What we're checking	Points
Agentic depth and MCP craft	Does the agent genuinely plan, call tools, and recover — or is it one model call in a loop? Are the tool boundaries ones a stranger could reuse?	30
Open-source rigour	Can we clone it and run it? Is the README honest about what is unfinished?	20
Fit to the farming cycle	Does a named office or collective do a real thing faster? Did you talk to an extension officer or a farmer?	20
Evaluation and reliability	Is the test set honest rather than curated to pass? Is run-to-run variation measured or hidden?	15
Defensibility	Could this recommendation survive a farmer asking "why"? Sourced findings, readable log, known cost per run.	10
Demo	Three clear minutes showing the agent working, including where it fails.	5
One hundred points in total.		100

Judging happens in two stages: a written review against the criteria above, shortlisting the top scores; then a 20-minute live technical review — ten minutes walking us through your code, then a new requirement added live while we watch.

Judges: panel to be announced.

Resources

There's no starter pack for this track. By this stage we're assuming you've already got MCP servers you trust and know how to stand up quickly — what data you build on is entirely your call: regional rainfall and vegetation data (CHIRPS, TAMSAT), FAO's GIEWS and WFP's price-monitoring bulletins, a national open-data portal, your own synthetic household set, or data you have consent to use from a real cooperative. Just be honest in your README about where it came from and what it doesn't cover.

If your agent's advisory only holds up because the underlying data is unusually clean, say so — real household and plot records rarely are, and criterion one rewards handling that mess, not assuming it away.

Questions any time in the community channel; every answer is posted publicly so no team gets a private advantage. Shortlisted teams move into a mentor-supported phase with access to better compute and, where relevant, data partnerships to take the prototype toward something a local agricultural office could actually pilot.

Rules

Eligibility: Open to developers aged 18 and over, resident anywhere in Africa. Solo entries or teams of up to three. Organisers, judges, and their immediate families may not enter.
Submission period: Opens TBC, closes TBC. Repositories are cloned at the deadline timestamp; commits after it are visible and will disqualify a submission.
Multiple submissions: One project per team. You may not enter the same project under more than one theme.
Ownership: You keep everything you build. You grant us permission to reference and demonstrate your project in community materials. Your repository must carry an OSI-approved licence to be eligible.
Data conduct: Whatever you build on is your choice — open data, your own synthetic set, or real records you have explicit consent to use. Never a real household's records without consent, and never anyone's personal details. Aggregate market and weather data are fair game.
AI assistance: Expected and not penalised — use whatever coding tools you normally use. The live review is where we separate directing a coding agent from accepting output you cannot explain.
Language: Submissions in English, wherever you are building from. Your agent may handle documents in any language.

Instant disqualification: A chatbot wrapper with no autonomous tool use · an MCP server that proxies a single endpoint · an agent that sends advice, places an order, or issues a credit flag without officer approval · a recommendation against a household with no source · data from a real household without consent · a repository that does not run · results presented as real runs that never happened · someone else's project, undisclosed.

Schedule

Suggested build milestones — same shape for every track.

Day 1
Kickoff, team finalisation, and deep-dive into sector-specific problem statements and available open datasets.
Days 2–3
Ideation, architecture design, and initial prototyping. Milestone: "Paper Prototype" review with domain mentors.
Days 4–5
Core development, API/MCP integrations, and UX refinement. Milestone: midpoint "Stress Test" with end-user representatives.
Day 6
Final polish, documentation drafting, and open-source repository packaging.
Day 7
Final submission via the challenge portal; selection begins.

Ready to enter agriculture?

Solo entries or teams of up to 3. You can add your repository link later.

Apply for Agriculture

The Extension Backlog