Read-only debugging agents for production: a capability envelope you can audit

Read-Only Debugging Agents for Production: A Capability Envelope You Can Audit (2026)

An agent that can look at a production incident is useful. An agent that can also change production is a second incident waiting for a bad day. Most teams resolve that tension by choosing one extreme: hand the agent the same credentials an on-call engineer uses, or keep it out of the incident entirely and go back to kubectl and grep. The direct answer to the middle path is this: “read-only” is not one control, it is four, and they do not fail together. The agent’s tool selection is advisory. The other three — the credential or API scope the call travels over, the database transaction mode, and the runtime guard on evaluated expressions — are enforced somewhere outside the agent’s goodwill. A session is only as safe as the outermost enforced layer, and it is only auditable if the enforcement decision is recorded per call. ...

November 16, 2026 · 13 min · baeseokjae

Agent Traces That Don't Join: gen_ai.system vs gen_ai.provider.name and a Join-Key Contract for Mixed-Framework Fleets (2026)

A team runs a LangGraph service, an OpenAI Agents SDK worker, and a coding-agent CLI with telemetry switched on this quarter. Every producer is instrumented, every span arrives, and the per-agent cost dashboard still covers exactly one of the three. The others are not erroring; they are simply invisible to the group by everyone wrote, because the attribute that identifies a provider is spelled differently by each producer and, in one case, is not an attribute at all. The instrumentation guide already on this site, OpenTelemetry for AI agents, teaches how to emit those spans and predates the rename described below; this post is about what happens after the spans exist. The architecture question that led readers to the site is answered structurally in the multi-agent system design guide, which covers how to decompose the fleet but not what the telemetry contract between its members has to say. ...

October 31, 2026 · 10 min · baeseokjae
Diagnostic map distinguishing an absent graph, a query with no matches, a failed ingestion pipeline, and a stopped worker. No aggregate failure rate is inferred from these categories.

Cognee Empty Recall: Diagnose Missing Agent Memory

A memory layer fails quietly in one specific way that matters more than any benchmark score: the agent asks for context, receives nothing usable, and answers anyway. Nothing raises in the application. The model’s confidence is unchanged, the user gets a plausible answer with no memory behind it, and no alert fires because from the client’s point of view the call returned. Cognee is a graph-and-vector memory library for agents, and its maintainers publish dated fleet telemetry as public GitHub issues. That telemetry is unusual: it gives per-release, per-day denominators for exactly this failure, so the silence can be measured rather than described. This article maps each visible symptom to one cause, one probe, and one fix, and keeps every number tied to the release and date printed in the issue it came from. ...

October 29, 2026 · 14 min · baeseokjae
Frugal Tokens: Exploring Cost and Usage Across Coding Agents

Frugal Tokens: How to Track Coding Agent Cost and Usage Per Session

Frugal Tokens is a free, local, read-only web dashboard that shows exactly how many tokens each of your AI coding agent sessions consumed and what those tokens cost, across OpenCode, Claude Code, PI, Codex, and Cursor. It runs entirely on your machine, never mutates your agent databases, and surfaces hidden costs like prompt-cache misses that most developers never see. This guide walks you through installing it, pointing it at your agents, and reading the per-session explorer to cut your coding agent spend. ...

September 5, 2026 · 10 min · baeseokjae
Swarm: Local-First Control Plane Agent Watching Every Claude Code Session Live

Swarm: Local-First Control Plane Agent Watching Every Claude Code Session Live

A local-first control plane agent like Swarm watches every Claude Code session on your machine live so you can see each agent’s tool calls, reasoning, token spend, and cost from one dashboard instead of cycling through terminal tabs. It keeps a ledger of who holds which task and worktree, turns CLAUDE.md “never do X” prose into real permission decisions, and runs entirely offline with no account and no telemetry. What Is a Local-First Control Plane for Coding Agents? A control plane is the management layer that decides how the agents underneath it behave. In IT operations, the control plane and the data plane are separated: the data plane does the work, while the control plane watches, meters, and governs. ra3orblade/swarm applies that same separation to AI coding agents, with one crucial twist: the entire control plane runs on your own machine. ...

September 5, 2026 · 12 min · baeseokjae
MCP as an Observability Interface: Connecting AI Agents to Kernel Tracepoints

MCP as an Observability Interface: Connecting AI Agents to Kernel Tracepoints

MCP observability turns the Model Context Protocol into a two-way interface: AI agents don’t just call tools, they receive ground-truth telemetry from kernel tracepoints, eBPF programs, and kprobes. By exposing low-level system instrumentation through MCP servers, agents get a real-world model of the live system instead of hallucinated state — closing the observability gap that traditional APM leaves wide open. What Is MCP and Why It Needs Observability The Model Context Protocol (MCP) is an open standard that standardizes how AI agents discover and invoke tools, resources, and prompts. Instead of every agent building bespoke integrations with every service, MCP defines a common protocol: a host (the agent runtime) connects to MCP servers, which expose tools the model can call and resources it can read. ...

August 24, 2026 · 10 min · baeseokjae
OpsGraph: Evidence-First Agentic PostgreSQL Investigations

OpsGraph: Evidence-First Agentic PostgreSQL Investigations

An agentic PostgreSQL investigation is an AI-driven workflow that explores a database to answer operational questions, and OpsGraph makes it evidence-first: it discovers an approved read-only schema, validates every SELECT against the PostgreSQL AST and policy bounds, and records a tamper-evident audit chain so no conclusion can be asserted without opening its exact evidence. Instead of giving a model direct credentials and “asking it anything,” OpsGraph constrains the investigation to bounded, read-only, auditable queries that fail closed on unsafe SQL. ...

August 22, 2026 · 14 min · baeseokjae
Bar Observatory: deterministic local audit reports for Claude Code

Bar Observatory: Deterministic Local Audit Reports for Claude Code

Bar Observatory is an MIT-licensed Rust tool that records Claude Code sessions into a local SQLite file and renders that file into JSON, HTML and Markdown audit reports. The selling point is narrow and specific: the render path contains no model, so an unchanged store produces the same document every time, and the same store can be handed to a reviewer who renders it again. If your question is “what did the agent actually do, and can someone else re-derive that answer?”, this is the category of tool worth evaluating. If your question is “how many tokens did this cost?”, it is the wrong tool. ...

August 21, 2026 · 13 min · baeseokjae
Agent Trajectory Monitor: Watching Tool Calls and Token Spend in Real Time

Agent Trajectory Monitor: Watching Tool Calls and Token Spend in Real Time

An agent trajectory monitor watches the full sequence of steps, tool calls, decisions, and token spend an AI agent makes in real time — not just the final answer. Because AI agents consume 5-30x more tokens per task than standard chatbots, real-time trajectory monitoring has moved from a nice-to-have to a budget requirement. It lets you see whether your agent is doing the right work or just getting the right answer by chance. ...

August 20, 2026 · 11 min · baeseokjae
TencentCloud AgentObs SDK for DeepSeek Harness

TencentCloud AgentObs SDK for DeepSeek Harness: Zero-Collector GenAI Trace Observability

DeepSeek Harness observability usually means standing up an OpenTelemetry collector, a tracing backend, and code instrumentation. The TencentCloud AgentObs SDK for DSH eliminates all of that: it is a DeepSeek Harness plugin that observes the harness’s native session, agent-loop, LLM-stream, and tool lifecycles and uploads GenAI trace spans directly to Tencent Cloud Log Service (CLS) using Protobuf. There is no OTLP collector, no sidecar, and no extra service to deploy. If you already run agents on Tencent Cloud, this is the fastest path from “where did my agent spend its tokens?” to a dashboards-and-alerts answer. ...

August 20, 2026 · 11 min · baeseokjae