200the HTTP status of a response whose tool call was cut off by max_tokens; truncation is reported in the body, not as an error

Truncated tool call: a 200 OK that cut the arguments off mid-JSON

An agent tool call truncated by max_tokens arrives as HTTP 200 with incomplete JSON. Which stop reason to check on Claude and OpenAI, and how to retry.

Tooling6 minOct 15cited

72%of high-fault runs in which a retry-only agent repeated a customer action

A timeout retry repeated the action in 72% of runs. Here is the fix

Retrying a tool call after a timeout duplicated the side effect in up to 76% of simulated runs. How idempotency keys and verify-before-retry stop it.

Tooling8 minOct 8cited

0.1xstandard base input price to re-read a cached prefix, against 1x to resend it

Four places to put agent memory. Only two survive a restart.

Summary, note file, retrieval index and prompt cache solve different problems. Teams pick one, use it for everything, and pay for the mismatch every turn.

Tooling9 minSep 24cited

83.0%value accuracy on text, against near-perfect schema compliance

Your schema always validates. The values are right 83% of the time.

Constrained decoding made malformed tool arguments rare. It did nothing for wrong ones, and most pipelines still read a parsed object as a correct one.

Tooling8 minSep 22cited

25 ms → 5.06 squoted against measured

Agent sandboxes: vendors quote 25ms, benchmarks measure 0.27s to 5.06s

Vendor cold starts for agent sandboxes span 25ms to 150ms. An independent 2026 benchmark measured 0.27s to 5.06s. Which of the two should decide.

Tooling9 minSep 16cited

2 / 5platforms self-hostable

Agent observability platforms compared: licence, price, self-host

Langfuse, LangSmith, Braintrust, Phoenix and Helicone priced side by side. The deciding factor is usually where your traces are allowed to live.

Tooling9 minSep 15cited

4 modelsmodels, same decline

A million-token window moves the cliff. It does not remove it.

Accuracy falls as context grows, across every model tested. Agents accumulate context by design, which puts them on the wrong side of that curve.

Tooling8 minSep 12cited

9 / 11registries compromisable

Nine of eleven MCP registries were compromisable. What do you allow?

MCP hit 41% production adoption before it had a trust model. Scans find 30 to 82% of public servers flawed and 8.5% using OAuth.

Tooling8 minSep 12cited

3span types for agents

Token counts are a span attribute now. Cost per step is a query.

OpenTelemetry's GenAI conventions put token usage on every span. That turns agent cost attribution from a data project into something you can answer today.

Tooling7 minSep 12cited