AI Agent

An AI agent is a language model that runs in a loop instead of returning one answer. It picks an action, takes it through a tool, reads the result, and picks again until the goal is met or something stops it. The basic block is a model with tools, retrieval and memory, and an agent puts that block inside a loop where the model, not your code, chooses the next step. The idea goes back to the 2022 ReAct paper, which interleaved reasoning ("the test failed on a null check") with actions ("open that file and look"). Acting with reasoning beat acting alone, and looking things up cut the invented facts that reasoning alone produced. The think, act, observe cycle of the ReAct pattern is still the skeleton of every agent we build.

Model, tools, loop, stopping condition#

An agent has four parts. A model decides what to do next. Tools, reached through tool calling, let it act on the world and read ground truth back: run a query, edit a file, call an API. A loop runs think, act, observe, over and over. A stopping condition ends the run when the task is done, an error cannot be recovered, a person has to check something, or the step cap is reached.

The loop matters because it handles tasks whose steps you cannot write down in advance. Ask an agent to make a failing payment test pass and no fixed pipeline gets you there. It runs the test, reads the stack trace, opens the file the trace points to, finds a currency rounding bug, edits, reruns, and repeats until the test passes. Each tool result changes the next move, and a fixed sequence of model calls cannot react that way.

Workflow or agent#

Not everything that calls a model is an agent. A workflow runs models and tools along code paths you wrote: classify the ticket, then draft a reply, then check the tone. An agent lets the model choose its own path. Anthropic's five published workflow patterns, such as prompt chaining (one call's output feeds the next) and routing (a classifier picks which prompt runs), cover a large share of real systems while your code stays in control.

Use a workflow when you can list the steps, and an agent when you cannot. Invoice extraction is a workflow: the same five steps every time, and a buyer of that system wants it to behave the same way every time. "Find out why signups dropped 18% last week" needs an agent, because the answer could sit in analytics, a deploy log or a broken email template, and the path only shows itself as evidence comes in. Anthropic's guidance and OpenAI's agent guide both say to start with the simplest design that works, such as one well-built prompt, then retrieval or a fixed workflow, and to accept an open loop only when those fall short. We agree, for a reason of our own: a buyer can review a workflow's output by its shape, and an agent's output only by reading the run. An agent costs predictability, latency and tokens. Anthropic measured its agents using about four times the tokens of a chat interaction, and its multi-agent research system about fifteen times.

Why autonomy multiplies error#

A single model call that is right 95% of the time feels solid. Put that reliability inside a loop and the arithmetic turns against you, because success has to hold at every step. At 95% per step, a 5-step run finishes clean about 77% of the time, a 10-step run about 60%, and a 20-step run 36%. By 50 steps you are at roughly 8%.

The real picture is worse than those numbers, because agent errors are not independent coin flips. One wrong step changes the context for every later step: an agent that misreads a database schema at step 3 writes confident, well-formed, wrong queries from step 4 onward, so each error steers the steps after it. Reliability work on agents therefore spends less effort on making any one step smarter and more on making bad steps visible and recoverable. That means ground truth the agent can check itself against (does the test pass, does the page render), and evals that score whole runs rather than single answers.

Limits the model cannot change#

Agents that run in production work inside limits set outside the model:

• A maximum step count and a token or dollar budget, so a confused agent stops early instead of running all night. Set the cap from the step counts you see in testing, and have work that needs more checkpoint and resume.
• The fewest tools that can do the job, read-only by default, with credentials scoped to the task rather than the system.
• Contained writes. Coding agents work on branches, and anything that touches production data runs against a replica or behind a feature flag until it has a record.
• A person approves anything that cannot be undone, such as sending money, deleting records or emailing customers. Reversible actions can run on their own.
• A log of every step, tool call and result. You cannot debug a 20-step run you did not record, and the logged runs become eval cases.

How long agents run unattended, as of October 2026#

Coding agents matured first, because compilers and test suites tell the loop when it is wrong. The same property is why our DevOps agent runs every template through a validation loop before AWS sees it: an agent improves fastest where something outside the model can say no. Anthropic reported in September 2025 that Claude Sonnet 4.5 kept working on one multi-step task for more than 30 hours. Computer use then took agents past APIs to software that only has a screen.

METR measures how long a software task, timed on human experts, an agent can finish at a given success rate. Its January 2026 update put the doubling time since 2023 at 131 days, about four months. By its spring 2026 estimates, GPT-5.4 finished tasks of about 6 hours and Claude Opus 4.6 tasks of about 12 hours at 50% success, and Claude Mythos Preview went past 16 hours, where METR says its task suite stops measuring reliably. At 80% success Opus 4.6 fell to about 1.2 hours. Plan around the 80% figure, not the headline one, because a production run that fails half the time still needs a person on every run. METR stopped updating the page in September 2026, so treat these as a trend rather than a promise about your workload.

Under all of this, the Model Context Protocol became the common way to plug tools into agents. Anthropic donated it to the Agentic AI Foundation under the Linux Foundation in December 2025, and the 2026-07-28 revision removed sessions and the opening handshake so that servers can be stateless. Harder problems are increasingly split across multi-agent orchestration, where a lead agent breaks up the goal and hands parts to focused sub-agents.

Where our agents fail#

On the executive operations agent we built, putting every tool definition in the prompt worked until roughly thirty tools. Past that point the model began choosing plausible but wrong tools, and nothing in the output said so. We replaced the tool list with a registry the agent searches at each step, so it chooses among a handful of definitions, and the registry grew past a hundred tools. The same build has a critic that decides continue, retry or escalate after every observation, because without one a single bad observation carries through every later step and the agent reports success on work it never finished.

The failures we design for on that build are mostly not model failures. A login does not go through, a page has changed since the plan was made, or a tool returns a partial result. Retries are bounded, because an agent that retries forever against a broken vendor burns money and time while looking busy, and the run log compares the agent's claimed completion with the actual state of the target application. The DevOps agent that changes live AWS accounts goes further: it has no general shell, its purpose-built tools are its whole action surface, and a template that will not come clean after a bounded number of revisions is escalated as a request the agent cannot safely fulfil, instead of going back to the model. On both builds, what still fails is the account, the session or the vendor rather than the model, so we start a new agent's design from the list of what its environment will do to it, and tune prompts after that.