GitHub Copilot runs solo-company departments
João Pedro Silva Setas reports using eight GitHub Copilot custom agents for company operations, producing social content and tracking deadlines while exposing shared-memory corruption.
dev.to
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Use these approved company records to run an operations check. Inspect errors, overdue tasks and deadlines, consult bounded specialist roles and prepare a prioritized plan. Keep shared memory writes safe, record uncertainties and escalate strategy or consequential decisions for my review.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- João Pedro Silva Setas
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Setas reports 84+ tweets, five DEV articles and tracked deadlines after two months, alongside very low revenue and a shared-memory corruption incident.
- Human involvement
- Setas retained strategic decisions and approved Improver diffs. Only memory writes required no approval; protected sections depended on instructions rather than machine enforcement.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Define company roles, instructions and tools in Copilot custom agent files.
- Run an operations standup that checks errors, overdue work and memory.
- Delegate specialist tasks and use bounded consultations for cross-domain decisions.
João Pedro Silva Setas runs five software products as a solo founder. He built eight GitHub Copilot custom agents for operations, finance, marketing, accounting, legal review, strategy, technical decisions and system improvement. Each role lives in a Markdown agent file, with tool access and domain instructions.
A daily operations run checks Sentry, the sprint board, overdue recurring work and a shared knowledge graph. It then delegates work and produces a prioritized plan. Roles can consult one another when a task crosses domains, such as marketing claims needing legal review or pricing needing tax context. Call-chain tracking and a depth limit prevent endless consultations.
Setas reports more than 84 tweets and five DEV articles after two months, with marketing taking him less than an hour a week. He also reports no missed compliance deadlines, but just €6.09 in revenue against roughly €42 a month in hosting. Those figures describe his own business and do not establish that the agents made it profitable.
One useful failure was concurrent writes corrupting the JSONL knowledge graph. Setas forked the memory server to add serialization, atomic file writes and repair during loading. He retained strategic decisions and approved the Improver agent's proposed changes. Memory writes alone needed no approval. Protected instruction sections depended on agent compliance rather than machine enforcement. The case shows both the reported operational output and the infrastructure work required to keep several tool-using roles from damaging shared state.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: DEV Community · João Pedro Silva Setas
Frequently asked questions
How did the agents share context?
They used an MCP knowledge graph stored in JSONL, holding decisions, deadlines, products and lessons.
What broke during concurrent work?
Parallel memory writes corrupted JSONL data. Setas added a mutex, atomic writes and load-time repair in a forked memory server.
Did the business become profitable?
The account reports only €6.09 in revenue and about €42 a month in infrastructure at that point, so it does not establish profitability.
More details
Remaining steps
- Review output and repair shared memory after concurrent writes corrupt the store.
About this account
Copilot roles read tools, maintain memory, delegate work and publish scheduled content under domain instructions without a prompt for every subtask.
Firsthand source opened over HTTP and read on 2026-10-15. Publication date comes from the source page. Outcomes are author-reported and were not independently reproduced. Suggested prompt written by Agents In Use. All metrics self-reported; no independent business or compliance audit; existing Copilot subscription cost not treated as zero total cost.
Credit: João Pedro Silva Setas
Source published
Added to Agents In Use
Source checked
Tools and topics
GitHub Copilot custom agents · MCP memory server · Sentry · X · DEV Community · Task scheduler
Solo operations · Specialist agents · Shared memory
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.