Topics

Agents

Every article on Agents published on this blog, newest first: 9 in total, written by Ramón Chancay.

9 articles on Agents

Illustration of an agent's hard limits: the agent loop enclosed in a frame of checks—budget, command allowlist, path allowlist—with a kill switch outside and a log of every turn

AI31 Aug 202633 min read

The hard limits of an autonomous agent: what keeps it from disaster

The limits that separate an experiment from an agent you'd leave running: iteration and token caps treated as a budget, command and path allowlists, a kill switch, idempotency and observability.

Illustration of an agent harness as a chassis: the model's loop in the center and four instruments wired to it (files and a terminal, a context being compacted, subagents, a switch with a lock and a gate); the task enters from the left, the report leaves on the right, and a person approves from outside

AI7 Aug 202614 min read

What an agent harness is (and when not to use one): deepagents vs LangGraph

What an agent harness is, what deepagents adds on top of create_agent and LangGraph, its four layers (environment, context, delegation, steering), and the table for deciding when to drop down to LangGraph.

Illustration of the end-to-end system: a Jira ticket enters through a trigger, passes through a stateful queue, runs in an ephemeral workspace, and comes out as a pull request that returns to Jira as a comment

AI2 Jul 202629 min read

From a Jira ticket to a pull request: the end-to-end autonomous system

How I turned the agent into an autonomous system: a Jira ticket triggers it (webhook or polling), a queue with persistent state processes it exactly once, and it returns a PR and a Jira comment.

Illustration of an isolated agent: a separate working copy branches off the repository on its own branch, the loop works inside it, and the output is a pull request, not a merge

AI30 Jun 202619 min read

The agent inside a real repo: isolating tasks with git worktree and shipping a PR

How to put an agent loop to work on a real repository: isolate each task with git worktree, reproduce the bug with a test, run the whole suite, and ship a PR instead of a merge.

AI30 Jun 202621 min read

From generating files to using tools: a code agent's ReAct loop

The shift from generating the whole function to an agent that reads, searches, edits, and runs with tools: the ReAct loop and why the agent builds its own context instead of receiving it in the prompt.

Illustration of three levels of input for an agent: hand-written tests, examples the agent turns into tests, and the repo's existing suite as the evaluator

AI30 Jun 202624 min read

Who writes the agent's tests: three ways to define what's correct

Who writes an agent's tests defines what counts as correct. Three levels of input: spec and tests by hand, examples the agent turns into tests, or the repo's suite as the evaluator.

Execution sandbox illustration: the model's text enters an isolated enclosure with a timer, passes through four filters, and comes out as a classified observation

AI30 Jun 202621 min read

The sandbox in an agent loop: from the model's text to a reliable observation

The sandbox in an agent loop turns the model's text into a reliable observation: extract the code from the prose, check that it parses, run it with a timeout, and isolate the process.

Write-test-fix loop illustration: one node writes code, another tests it, and the failure feeds back as context until the tests pass

AI30 Jun 202624 min read

The simplest loop that works: a write-test-fix agent, step by step

The simplest agent that works: spec and tests by hand, the model writes the code, it runs, and the failure feeds back as context. A complete write-test-fix loop in code that runs.

Agent loop illustration: a goal enters a cycle of decide, act, and observe that repeats until a stopping condition

AI29 Jun 202612 min read

What an agent loop is (and isn't): state, action, stop

What an agent loop is and isn't: its three components—state, action, and stopping condition—and why 'build me an app' is input for a chat, not an agent.