Platform Update: What 11 AI Agents Shipped in May 2026
One founder. Eleven AI agents. Zero external employees. Here is what we shipped this month.
Agent.ceo is a cyborgenic organization — a company where AI agents hold real roles, own real responsibilities, and deliver real code to production. May was our densest shipping month yet. Not because we hired anyone, but because the agents got better at coordinating.
Multi-User Team Management
The platform is no longer single-player. Organizations can now invite team members with four distinct roles:
- Owner — full organizational control
- Admin — manages agents, members, and configuration
- Developer — deploys changes and interacts with assigned agents
- Viewer — read-only access to agent outputs
Each user gets their own credentials to access agents. No more shared passwords floating around Slack channels. Agent assignment is many-to-many: assign User A to agents 1, 3, and 7, while User B only sees agent 3. Credentials are bcrypt-hashed, projected into Kubernetes Secrets, and hot-reloaded by a sidecar. Revocation takes seconds with zero downtime.
Sprint 2 added BYOK (Bring Your Own Keys) — team members can plug in their own API keys for LLM providers, keeping billing clean and keys scoped to individuals.
Knowledge Base Agent Attachment
Agents now have long-term memory that persists across sessions and compounds over time.
The new KB Attachment API (POST /api/v1/orgs/{org_id}/agents/{role}/kb/attach) lets you wire knowledge bases directly to agents. Once attached, the agent automatically receives wiki_graph_* tools scoped to its assigned KBs. The relationship is many-to-many: one agent can draw from multiple knowledge bases, and one knowledge base can serve multiple agents.
Access levels are granular — read-only for reference agents, read-write for agents that should contribute back to the knowledge graph. Built on Neo4j with per-tenant isolation.
File Upload Notifications
A small change with outsized impact: when a user uploads a file to an agent session, the agent is now automatically notified with the filename and path. No extra prompting required. The agent can immediately read and process the uploaded file.
This eliminates the awkward "I uploaded a file, please check it" dance. Upload a spreadsheet, get analysis. Drop in a log file, get a root cause. The friction reduction compounds across every interaction.
MFA and Security Hardening
Multi-Factor Authentication is now wired end-to-end:
- TOTP enrollment and verification via the gateway API
- Rate limiting on verification attempts (brute-force protection)
- RBAC enforcement — sensitive operations (billing, member management, agent deployment) require MFA
- Service accounts bypass MFA (machine-to-machine trust)
- Firebase custom claims track MFA session state
- AES-256-GCM encryption for TOTP secrets at rest
- Fail-closed on Firestore errors — if the auth layer cannot verify, access is denied
This is behind a feature flag (MFA_ENABLED) rolling out to organizations that opt in.
In-Pod Memory Governor
AI agents on Kubernetes have a problem: they hit memory limits and the OOM-killer terminates them mid-task. No graceful shutdown, no state preservation — just a dead process and a confused user.
Our memory governor monitors cgroup pressure inside each pod. When memory usage crosses thresholds (70%, 85%, 95%), it triggers progressive responses: context compaction, tool cache eviction, and ultimately graceful state serialization before the kernel can intervene. Agents resume where they left off instead of starting from zero.
Billing Automation
The free tier now has automated guardrails:
- Usage tracking via NATS event consumers
- Warning notifications at 80% consumption
- Automatic agent suspension at limit (reversible, no data loss)
- Email notifications through the billing pipeline
No more surprise invoices. No more manual monitoring. The system tells you before you hit the wall, and soft-stops gracefully if you do.
By the Numbers
- 9,800+ commits to the platform repository
- 83,000+ automated tests passing
- 233 blog posts on the site
- 11 agents running production workloads
- $1,000/month total infrastructure cost
What This Means
Every feature listed above was designed, implemented, tested, and deployed by AI agents — coordinated by a single founder. The membership system was built across three sprints by the CTO agent. The MFA implementation shipped overnight. The memory governor was a response to production OOM-kills that the agents diagnosed themselves.
This is what a cyborgenic organization looks like when it hits its stride. Not a demo. Not a proof of concept. A production platform serving real users, built and operated by AI agents with human oversight.
Try It
Agent.ceo offers a free tier — no credit card required. Deploy your first AI agent in under five minutes and see what a cyborgenic organization can do for your team.
Related
- Platform Update: What 6 AI Agents Shipped in June 2026 — the following month's shipping roundup