AI
Git as Agent Memory
MintMCP created a Git repository for business operations, not code. The first commit set up a weekly customer update system. Each week a new markdown file was added. The agent used Claude code with a Claude MD file to generate updates. Over time the files accumulated and became agent memory. Git provides auditability, revert, and multi-agent merge. Agents can update their own Claude MD configs, enabling self-learning. This pattern uses GitHub for memory and MCPs for external services like Linear and GitHub itself.
Tracking Work with Progress Files
Agents need to know where they left off. MintMCP added a CSV append-only log and a progress.md file. Claude MD instructions tell the agent to update these files on every action. On session start the agent loads the last 100 lines to resume work. Dates remain a weakness; agents must use bash date commands to get the current day right. Each agent now has its own folder structure. The deal agent has a deals folder, the social agent has people and company folders.
MCP Gateways for Security
Background agents raise security concerns: prompt injection, fine-grained permissions, OAuth complexity. MintMCP introduced an MCP gateway as an abstraction layer between agents and backend services. The gateway groups tools by use case and role. A sales agent may read CRM records and drop emails but cannot delete records or send emails until trust is established. The gateway manages authentication (OAuth, bearer keys) and issues short-lived tokens. This approach also mitigates the “little tri-factor” — untrusted content, private data access, and external communication — by tightly bounding tool access.
Slack as Universal Agent Interface
MintMCP moved all agents into Slack. Every thread runs in an ephemeral sandbox. The engineering bot (Dan Loop) follows a plan-implement-review loop: Claude writes code, Codex reviews via MCP, Claude fixes, and the cycle converges. The sandbox uses Vercel. When done, Slack notifies the human. Developers no longer watch terminals. Anyone in the organization can request a bug fix by messaging Slack. An agent can fork a repo, implement, and submit a pull request. All agents — engineering, sales, procurement, even a food bot — now live in Slack and can be woken by cron jobs.
Hooks and Agent Monitoring
Background agents need observability. MintMCP uses hooks from Claude code’s agentic loop: pre-tool use, post-tool use, task created, task completed. Each hook triggers a script that logs telemetry and applies guardrails. The live dashboard shows every tool call, command, and action for each agent. Dan Loop’s activity is visible in real time. With agents waking at 6 a.m., monitoring ensures nothing goes wrong without visibility.
Q&A
How does MintMCP identify and define each agent? Each agent gets a separate Slack app with its own API keys, delegated credentials, and permissions via the MCP gateway. ▶ Watch (22:17)
How do you manage a federation of agents without manual Slack app creation? Agents are created programmatically from Slack manifests. The number of agents grows as distinct functions; too many functions in one agent makes evaluation hard. ▶ Watch (23:11)
What happens when two developers ask Dan Loop for code changes at the same time? Each Slack message creates a separate ephemeral sandbox with a cloned repo. Agents never commit directly to main; they always submit pull requests for human review. Merge conflicts are resolved in the pull request. ▶ Watch (24:58)
Notable Quotes
Our human-to-agent ratio is about 1:5 right now. Jiquan Ngiam · ▶ Watch (1:28)
we were able to merge that pull request even before the call ended. Jiquan Ngiam · ▶ Watch (3:08)
Git is fully auditable. You can go back in time, get blame, see when the model did it. Jiquan Ngiam · ▶ Watch (6:12)
We ran Codex as a MCP server for Claude. And so Claude actually calls Codex via MCP. So everything actually runs in Claude. I like to joke that Claude is bossing Codex around. Jiquan Ngiam · ▶ Watch (15:21)
The more you can bring your agents into Slack, the more you would be able to leapfrog the AI adoption curve because everyone in the organization knows how to communicate on Slack. Jiquan Ngiam · ▶ Watch (21:24)
Key Takeaways
- Git repositories work as auditable, multi-agent memory stores for autonomous agents.
- Progress tracking files with Claude MD instructions enable agents to resume work across sessions.
- MCP gateways provide fine-grained tool permissions and mitigate prompt injection risks.
About the Speaker(s)
Jiquan Ngiam was a senior staff researcher at Google Brain and founding team member at Coursera, where he helped build Andrew Ng’s online machine learning course from the ground up. He co-authored pioneering work in multimodal deep learning at Stanford. Currently co-founder and…