Reinventing.AI

AI Agent Trends

AI Agents Trends: Model Retirements, a Duplicate Write Bug and Fallback Routing Put Agent Reliability in Focus

GitHub retired three Copilot models on November 16, a developer reported approved MCP writes running twice, and LangChain raised $300M for agent monitoring.

AI Agent Insights Team6 min
AI Agent Production Reliability

Three developments in the past week show where agent workflows break once they run every day. GitHub retired three models from Copilot on November 16 and scheduled six more for December 14. A developer on the OpenAI community forum documented a case in which a single approved tool write ran twice. And LangChain raised $300 million on November 13 for LangSmith, its platform for watching, scoring and running agents in production.

None of the three is a new model or a new agent. Each one concerns what happens after an agent is set up: the model underneath it changes, a tool call repeats, or nobody can see what the agent did. For solo operators and small teams that run agents on a schedule, these are the failures that cost money or data without an error message.

Models are retired on a schedule

GitHub's November 16 changelog lists three models deprecated that day across Copilot Chat, inline edits, ask and agent modes, and code completions: Claude Opus 4.8 and Claude Opus 5, with Claude Opus 5.5 as the suggested replacement, and GPT-5.6 Terra, with GPT-6.1 Sol or GPT-6 Luna suggested instead. GitHub first announced these retirements on October 15, which gave users one month of notice.

Three days earlier, a second changelog entry named the next group, all to be removed on December 14, 2026:

  • Claude Fable 5, replaced by Claude Fable 5.2.
  • Claude Haiku 4.5, replaced by Claude Haiku 5.5, except for individual subscribers on annual plans.
  • Gemini 3.8 Flash, replaced by Gemini 4 Hydrogen.
  • Gemini 4 Argon, replaced by Gemini 4 Carbon.
  • Grok 4.6, replaced by Grok 4.7.
  • Kimi K3, with the replacement listed as to be announced.

Both entries ask users to update their workflows and integrations to supported models. For Copilot Business and Enterprise accounts, the replacements are switched on automatically unless an administrator has turned off default model enablement. A workflow that names a model explicitly, rather than taking the default, does not move by itself. The replacement is also a different model, so an agent that passed its checks on the old one can behave differently on the new one.

An approved write that ran twice

On November 16 a developer filed a bug report in the OpenAI Developer Community titled “Responses API: approved native MCP write executes twice after a programmatic read.” The setup was narrow. A write tool on a Model Context Protocol server required human approval. After the approval was sent back, the same request went on to make a read through OpenAI's hosted programmatic tool calling.

According to the report, the server logs showed the write, then the read, then the write again. The two write records carried the same approval ID and the same arguments, and the test produced two changes on the server. The model's final message stated that the write had been called once. Control tests without the programmatic read produced one write. In a follow-up on November 17, the reporter found that resending the approval after a lost response raised the count to three writes in the programmatic read case.

An account named OpenAI_Support replied on November 16 that the company was looking into the issue. The thread shows no fix or root cause as of this writing. A community reply listed three workarounds: send the approval in one request and the read in a separate one, give the write tool a required idempotency key so the server returns the stored result for a repeated key, and check the response for two tool calls with the same approval ID instead of trusting the model's own summary.

The approval step did its job here. A person approved one action. The duplication happened after the approval, which is why a human checkpoint alone does not protect a write that charges a card, sends an email or edits a record. Earlier coverage of approval prompts and leaked secrets reached a related point about what approvals can and cannot catch.

Fallback routing for outages

The third failure is the simplest: the model is unavailable. AWS added model fallback to Amazon Bedrock AgentCore Gateway in October. The announcement says the gateway sends a request to another model or provider when the first is throttled, out of capacity or unreachable, and that a circuit breaker skips destinations that are currently failing.

The AgentCore documentation sets the limits. A rule can list one or two alternative destinations. Fallback triggers on chosen HTTP error codes such as 429 or 5XX, on connection errors, or on a timeout before the first byte arrives. Successful responses never trigger it. Once a streamed response has started, no fallback is possible, and the caller receives the partial response and an error. Each outbound attempt is billed as a gateway call.

Fallback keeps an agent running during an outage. It also means the agent may finish the job on a model it was not tested on, which connects back to the retirement problem above. Readers who run a single model can find the cost side of this choice in earlier coverage of routing between cheap and frontier models.

Money for watching agents

LangChain announced a $300 million Series C at a $3 billion valuation on November 13, led by CapitalG. The company says more than 7,000 organizations use LangSmith, and it describes the product as a platform to run, improve and govern agents. Its listed parts include tracing of what agents do, evaluation that scores agent performance, and gateways that manage model budgets and fallbacks. The post says Clay uses LangSmith to monitor hundreds of millions of agent runs per month. LangChain plans to show new LangSmith features on November 20.

The size of the round is a signal that investors expect agent monitoring to be a lasting need. The underlying practice does not depend on any one vendor. A trace of each run and a fixed set of test tasks are what make the two problems above visible.

Practical steps

The following steps are general guidance drawn from the sources above.

List every model an agent names. Compare that list with the provider's deprecation notices. GitHub's next removal date is December 14.

Rerun tests before switching. Run the same tasks on the replacement model and compare results. The site's guide to evaluating an agent workflow describes how, and earlier reporting on eval loops for solo operators covers a weekly routine.

Make write tools safe to repeat. An idempotency key on any tool that changes data turns a repeated call into a lookup instead of a second change.

Count actions from logs. The OpenAI report shows a model summary that disagreed with the server. The server's own record is the one to trust.

Test the fallback model too. If a gateway can route to a second model, that model needs to pass the same tests. Readers new to the category can start with the agent fundamentals.

What to watch

GitHub's December 14 removals, OpenAI's response to the duplicate write report, and the LangSmith features due on November 20 are the next dated items. Each will show how much of agent reliability providers handle and how much stays with the operator.

Sources