Building autonomous AI agents. Notes from the team building Gravity. AI workflows, the future of recurring work, and what we learn along the way.
An agent that can put events on your calendar saves an hour a week. An agent that can put events on the wrong calendar costs you a customer. The difference is two layers of allow-listing, an explicit timezone…
Read post →
A human approval step is the cheapest insurance policy in agent operations and the most over-applied governance pattern. Done well it catches the actions that should never happen and stays out of the way for…
Read post →
"Is this an agent or just an LLM?" is the most useful question a buyer can ask in 2026, and the one most vendors avoid answering directly. The distinction is not pedantic. An LLM is a component; an agent is a system…
Read post →
The first time an agent does the wrong thing in production is the day a trust model becomes a budget line. Every team eventually writes one. The question is whether you write it before the incident or after. This…
Read post →
Safety for AI agents is structurally different from safety for chatbots. A chatbot that says something inappropriate creates a screenshot. An agent that does something inappropriate creates an incident: an email…
Read post →
Eight months into 2026 and the agent market has visibly matured in three ways the 2025 forecasts mostly missed: pricing structures are converging, buyer procurement bars have hardened, and the marketing language of…
Read post →
"Why did the agent fail?" is the question every operator asks the first time an agent misses. The honest answer is almost always one of eight things, and the eight things are different enough that lumping them…
Read post →
"Is the agent any good?" is the question every buyer asks and almost no buyer can answer with a number. The shortage of good answers is not because the metrics are unknown; it is because most vendors publish one or…
Read post →
The benchmark landscape for AI agents in 2026 is busier than the buyer landscape can absorb. Five benchmarks dominate the conversation: GAIA, SWE-bench, AgentBench, BFCL, and ToolBench. Each measures something…
Read post →
The shift from static RAG to agentic RAG is one of the more useful generalisations in the AI agent stack, partly because it clarifies what the agent is doing and partly because it produces measurably better answers…
Read post →
The first time someone writes an agent prompt the way they write an LLM prompt, the agent breaks within the first hour of running. Not because the prompt is wrong in a literal sense; it is just shaped for the wrong…
Read post →
Most AI agents stop after one task. They run the first step, return a confident-sounding output, and then either silently halt, hand back to the human, or hallucinate a "task complete" status that does not match…
Read post →
Processes change. New CRM field, new approver, new review step, new tool. The agent that was right last quarter is silently wrong this quarter, and the gap shows up as runs that look fine on the surface but produce…
Read post →
An AI agent that has not been tested is an AI agent waiting to do something embarrassing or expensive on your behalf. Testing an agent looks different from testing software because the agent does not have a fixed…
Read post →
Sharing an AI agent with a team is the moment most agents quietly turn into a liability. The agent that one person built, supervised, and trusted now runs on inputs from people who did not write the prompt, with…
Read post →
An AI agent without a spending cap is an open tab on a model provider. Most of the time the bill is small. The expensive day is the one where the agent loops on a malformed input, or chains a search tool with itself…
Read post →
Restricting an AI agent to business hours is one of the cheapest reliability wins available. Most agent incidents are not catastrophic; they are awkward. An automated follow-up arriving at 3 a.m. looks like spam. A…
Read post →
Build vs buy is the wrong opening question for AI agents. The right opening question is: what would have to be true for either answer to be obvious. The four-axis framework that follows (cost, time, capability,…
Read post →
Autonomous AI and assistive AI are usually discussed as if they are different products. They are not. They are different points on the same spectrum, and the spectrum has five measurable axes: decision-making,…
Read post →
The Monday morning KPI summary is the report that should be automated and almost never is. The data exists. The query exists. The template exists. What is missing is the half-hour every Monday that somebody spends…
Read post →
Tool use is what separates a chatbot from an agent. A chatbot talks about sending the email; an agent calls the email-send tool and watches for the result. The mechanism under tool use is function calling,…
Read post →
Whether AI agents "reason" is a debate that often misses the practical point. The practical point is that different reasoning patterns produce different reliability characteristics on different tasks.…
Read post →
Orchestration is the runtime layer that coordinates multi-step agent execution. The LLM thinks; the orchestration decides which step runs next, retries when something fails, evaluates whether the goal is met, and…
Read post →
The discourse around AI agents in 2026 carries a lot of myths. Some come from vendor marketing; some come from social-media hot takes; a few are honest misunderstandings of fast-moving terminology. This post takes…
Read post →
AI agent memory is not one thing. It is three layers, each handling a different timescale and a different question. Short-term memory holds what is happening right now. Long-term memory holds what the agent might…
Read post →
Procurement conversations about AI agents fail when buyer and vendor use the same words to mean different things. This glossary defines 28 terms that show up in agent procurement, organised by category. Each entry…
Read post →
The post-meeting half hour is the most common place where good intent dies. People agreed to do things; nobody captured who, by when, or what exactly. The follow-up email never goes out. The action items never become…
Read post →
An inbox triage agent is the most popular first agent for a reason. The job is well-defined (read inbox, produce a summary), the failure mode is mild (a wrong summary, not a wrong send), and the value is immediate…
Read post →
Competitor tracking is the use case where the agent shape really pays off. The work is repetitive (read public pages on a schedule), the inputs are stable (a known list of URLs and accounts), the failure mode is mild…
Read post →
Cold lead follow-up is the use case sales teams want most and the use case where an AI agent is most likely to misbehave. The mechanics are easy: read a lead record, compose a follow-up, send. The hard part is…
Read post →