aiagent.

Validate Ollama Tool Calls with Pydantic and Retry Malformed JSON

Validating local Ollama tool-call arguments with Pydantic models before dispatch, plus a retry-on-invalid-JSON loop that tolerates smaller local models emitting malformed tool calls under load.

Claude Code split and merge subagent batched refactor

Decomposing a 50-file refactor into 10 parallel subagents via the split-and-merge pattern, with an explicit merge-time JSON schema and concrete throughput numbers compared to the sequential baseline.

claude-code-mcp-sigterm-keepalive-pattern

MCP Tool OAuth: The WWW-Authenticate Metadata

Observability for Claude agents: tracing, cost tracking, and the dashboard you actually need

The fifth article closes the series with the production-readiness layer. Cover three observability planes: (1) per-run trace — every message, tool call, tool result, token count, and latency — emitted as structured JSON logs, (2) cost tracking — input vs output tokens per run, daily rollup, per-tool cost attribution, (3) the minimal dashboard — three panels (success rate, p50/p95 latency, daily cost) built with Grafana + Loki or a 50-line FastAPI + HTMX page. Unique angle: skip the 'use LangSmith / use Arize' recommendation and show a free, self-hosted stack that works for solo operators. Reader gets a `TracingAgentLoop` decorator that emits OpenTelemetry spans, a cost-calculation helper keyed to current Claude pricing, and a screenshot-walkthrough of the three-panel dashboard.

Error recovery and the tool-budget loop: stopping agents that spiral into 200 tool calls

The fourth article addresses the failure mode that kills production agents: infinite retry loops, tool-call spirals, and silent context exhaustion. Cover four concrete safeguards: (1) per-task tool budget with hard cutoff, (2) repeat-call detection (same tool + same args twice in a row → force a reflection step), (3) error classification (transient vs permanent → retry vs escalate), (4) explicit 'I'm stuck' tool that lets the agent ask for human input rather than guess. Unique angle: instrument a real agent against a deliberately broken tool (flaky network, malformed JSON, rate-limit 429s) and show the difference in tool-call count and final success rate with vs without each safeguard. Reader gets a `BudgetedAgentLoop` wrapper, the repeat-detection logic, and a runnable test harness that injects failures.

Step 1: Scaffold a minimal MCP server with SSE transport and one slow tool

When Your Local AI Agent Forgets: The Silent 1024-Token Context Ceiling in Ollama

mcp resources/updated notification client fallback

Design an MCP server that emits `resources/updated` per spec but degrades gracefully on clients that currently ignore the notification (Claude Desktop, Claude Code) — etag-on-tool-call + lazy revalidation as a polling fallback so the same server works in both worlds.

MCP Tool Error Response vs Python Exception: When to Raise and When to Return

Two error paths exist in the MCP python-sdk. Picking the wrong one breaks the LLM's retry loop or crashes your session. Here's the rule, plus the pattern that survives SDK version drift.

PreToolUse Hooks in Claude Code: Validating Bash Before It Runs

Build a PreToolUse hook for Claude Code that vets every Bash command, blocks dangerous patterns, and logs the rest. Includes a working Python validator and denylist patterns.

Isolate the Anthropic SDK Behind One Adapter: A Pattern for Agent Codebases

Wrap Anthropic Sonnet calls in a single adapter module to get testability, prompt versioning, retry policy, and cost tracking without scattering SDK calls across your agent codebase.