Building autonomous AI agents. Notes from the team building Gravity. AI workflows, the future of recurring work, and what we learn along the way.
I have pitched ideas that were technically better than the thing that got approved, and lost. The lesson stuck: the quality of an agent project rarely decides whether it gets funded. The quality of the buy-in does. A…
Read post →
Most teams do not migrate to AI agents because a vendor sold them on it. They migrate because a Zap broke on an edge case for the third time, or because a Make scenario grew into a 22-step chain that nobody dares…
Read post →
"How long will this take?" is the first question every buyer asks and the one most vendors answer badly. The honest answer is that it depends on which of four very different things you are actually building. A…
Read post →
An executive does not read a business case to learn. They read it to decide whether to bet a slice of budget and reputation on you being right. Everything in the document either reduces their uncertainty or wastes…
Read post →
Choosing an AI agent platform looks easy until you have to defend the decision to a CFO, a security review, and a procurement officer simultaneously. The marketing pages are interchangeable. The pricing is opaque…
Read post →
Agent updates fail badly when they are in-place. A prompt edit lands mid-run and the second half of a conversation no longer matches the first. A new model lands and a tool-call signature shifts. An index rebuild…
Read post →
ROI calculations for AI agents fall into two categories: defensible numbers backed by measurement, and made-up numbers backed by vendor claims. The CFO can tell the difference. This guide is the defensible version:…
Read post →
An agent PoC succeeds when the go-no-go decision is obvious within 6 weeks. It fails when scope creeps, baseline is missing, or no one is responsible for the call. The 25-item checklist below covers what to confirm…
Read post →
This is the RFP template I send when buyers ask "give me the question list". Sixty questions across six sections, with a 1-to-5 scoring rubric and walk-away criteria. The template assumes enterprise procurement;…
Read post →
A PoC tells you the technology works. A pilot tells you the deployment works. Most teams skip the pilot because the PoC succeeded; then production hits real volume, real users, and real operational concerns, and the…
Read post →
Most agent platform outages I have seen were not catastrophic. A model provider had an incident; a region's vector store throttled; a deploy clobbered a prompt store; a tenant's run history was deleted by a buggy…
Read post →
Both Anthropic and OpenAI now ship enterprise chat surfaces that look almost identical in the screenshot: a chat thread, file uploads, custom instructions, connectors, and the ability to schedule or repeat work. The…
Read post →
The phrase "AI agent" now covers two products that look similar in a screenshot and behave very differently in production. The first is a chat-first workspace agent: ChatGPT with connectors, files, code interpreter,…
Read post →
SOC 2 is the buyer-facing artifact most enterprise prospects ask for before they let an AI agent platform near their data. It is also one of the most misunderstood. The report does not certify your AI; it attests…
Read post →
Most AI agent purchases go wrong at the sales-call stage, not after deployment. The team likes the demo, the vendor likes the deal, and a year later someone is paying for an unused seat tier with a 60-day notice…
Read post →
The classic SaaS isolation problem is well understood: keep tenant data, queries, and identity separated through the request path. An agent platform adds two new surfaces that have to follow the same rules. The…
Read post →
Agent logs grow fast. A single run easily writes dozens of structured events: orchestrator steps, model calls with input and output bodies, tool calls with payloads, retrieval queries with chunk text. Multiply by…
Read post →
The point of a canary is to learn things evals cannot. Evals run on a held-out set; production runs on whatever showed up today. Some regressions are visible only at production scale, on production traffic shapes,…
Read post →
Fuel accounts for roughly 24 percent of total fleet operating costs, according to the American Trucking Associations (ATA, 2024). That's the single largest controllable expense for most carriers. And yet the average…
Read post →
Construction is a $1.36 trillion industry in the United States alone (U.S. Census Bureau, 2025). Yet it remains one of the least digitized sectors on the planet. Projects run over budget 80% of the time. Rework eats…
Read post →
Picking the wrong AI agent vendor costs more than the subscription fee. Across The Standish Group's CHAOS research, software projects have never had a majority success rate: recent CHAOS data puts roughly 31% of…
Read post →
Most teams deploy AI agents and then track nothing. Or they track one metric, usually accuracy, and call it done. That approach misses most of the picture. Gartner predicts over 40% of agentic AI projects will be…
Read post →
An AI agent that calls five APIs holds five sets of credentials that an attacker can steal. That's not a hypothetical risk. The 2024 IBM Cost of a Data Breach Report found that stolen or compromised credentials…
Read post →
Your AI agent works. It answers questions, calls tools, returns useful output. But it takes eight seconds to respond, and your token bill keeps climbing. Sound familiar? Performance tuning is the difference between…
Read post →
Your company runs 15 agents across four departments. The LLM bill arrives as a single line item. Finance asks: "Who spent what?" You don't have an answer. That gap between aggregate spend and per-team accountability…
Read post →
Prompt injection is the #1 risk on the OWASP LLM Top 10 (OWASP, 2025). Agents amplify every LLM risk by adding tools, persistence, and autonomy. This checklist gives you 47 controls across 10 categories. Each control…
Read post →
A handoff is the contract between two agents (or one agent and a human) that specifies what gets passed, when, and what happens if the receiver is unavailable. Eight patterns cover most production cases. LangGraph,…
Read post →
Naive retries amplify outages; smart retries absorb them. The Google SRE book defines a retry budget so retries can't exceed a fixed fraction of normal load (Google SRE, ch. 22). For AI agents the same logic applies…
Read post →
You can build an AI agent in a weekend. Finding people who will pay to run it is the hard part. Most builders default to GitHub, get a few stars, and call it shipped. Three months later, the repo is collecting dust…
Read post →
Builders ask the same question on day one: "How do I get my agent to the top of Gravity?" The honest answer is that we do not sell that spot. Not to the loudest builder, not to the highest bidder, not to ourselves.…
Read post →