Notes from the project

2026

• Sep 24You can't roll back an agent5 minRollback undoes code, not consequences. An agent's value is acting outside your systems, so recovery has to be designed in before the action, not after.
• Sep 17The agent audit trail auditors will ask for — and why most teams don't have one6 minRegulators converge on one demand — prove who approved what, human and agent. What that log must contain, where agent platforms put the evidence instead, and the honest limits.
• Sep 10Updating a running agent is a migration, not a deploy5 minTeams redeploy agents like stateless functions, then wonder why a "small change" silently corrupts weeks of working memory. A stateful agent update is a database migration — and most teams have no migration tooling at all.
• Sep 3Prompt injection can't be solved. It has to be contained.5 minThe industry has stopped pretending it can filter injection away. The ones that survive treat it as a blast-radius problem — and bound it with permissions, not prompts.
• Aug 27Orphaned AI agents are the incident you haven't budgeted for5 minAgent fleets double quarterly, but only a fifth of teams decommission agents. Orphaned agents keep credentials, memory, and budgets alive long after their owners are gone.
• Aug 20Agent authorization is the gap that identity misses5 minWho an agent is matters less than what it's allowed to do. Frameworks ship auth-n but leave the scoping that actually contains blast radius to you.
• Aug 16Local-first agents — what "running AI on your own infrastructure" really means5 minSelf-hosting agents isn't about refusing the cloud. It's a thin-waist play: keep the sensitive reasoning on your infra, route only what needs it out. What that takes, and where it breaks.
• Aug 6Agent memory is an operations problem, not a model problem4 minEvery agent pipeline stores memory as a vector insert. The expensive failures happen after that — ownership, traces, consistency, and deletion nobody can prove.
• Jul 30Agent observability starts at the platform, not the patch6 minAgent observability is a $2.2B market in 2026, but most teams bolt it on after the fact. The real leverage is a platform that ships it as a default.
• Jul 23MCP won the standards war. Now the ops layer is the gap.4 minMCP is the default protocol for AI tool access, but its security and ops holes are where incidents happen. The protocol is a data format — not an operations layer.
• Jul 16Switching AI agent frameworks costs $315K — and most teams don't know they've locked in4 minFramework lock-in is the hidden cost of AI agents. Breaking-change cycles, provider price hikes, and migration bills averaging $315K are the real barrier to production.
• Jul 9Your multi-agent pipeline has no message authentication5 minEvery major agent framework passes messages between agents as plaintext — no signatures, no verification. That is the defining security gap of multi-agent systems in 2026.
• Jul 2AI agent costs blow up because the infrastructure has no budget layer4 minA $47k LangChain loop, Uber's annual AI budget gone in four months, and why per-agent budget enforcement beats monitoring every time.
• Jun 25Why 88% of AI agents never reach production4 minMost teams test agents by eyeballing outputs. The 67/10 gap — 67% see gains, 10% reach production — is an evaluation problem, not a model problem.
• Jun 19Cost control is the missing infrastructure for AI agents7 minA market-research pipeline of four agents entered a loop that ran for 11 days and cost $47,000. No framework prevents this at the infrastructure layer.
• Jun 19Why multi-agent coordination is harder than you think4 minThe hard part of multi-agent systems isn't building agents — it's how they coordinate. Most frameworks leave that as your problem. Workgroups shouldn't be an afterthought.
• Jun 15The 90% gap in AI agents5 min90% of agent projects stall before production. Models aren't the bottleneck — the ops layer is. And most teams build it from scratch.
• Jun 11Agents Don't Know Who They're Talking To6 minEvery major agent framework identifies agents by a name and a role string — no keys, no signatures, no peer verification. What production actually needs.
• Jun 11Running agents as an organization, not a developer toy4 minMost agent tooling was built for developers wiring up demos. Operating persistent agents as a company is a different problem — identity, budgets, governance, and trust by default.
• Jun 4Alpi vs the agent frameworks — a fair comparison5 minWhere Alpi sits next to LangGraph, CrewAI, and the vendor SDKs. Frameworks build agent flows; SDKs use one model; Alpi operates persistent local agents as an organization.
• May 28What Alpi costs — one metric, public bands, no metering4 minAlpi's pricing in full: a single billing unit (the installed agent), public graduated bands, grandfathered entry prices, founding-customer terms, and implementation services.