---
type: "post"
id: "2026-in-llms-so-far-analysis-fc134d"
slug: "2026-in-llms-so-far-analysis-fc134d"
title: "2026 in LLMs (so far)"
description: "Thesis: Coding agents appear to have crossed a practical usability threshold, shifting the limiting problem from code generation alone toward controlled deployment."
retrieval_nugget: "Thesis: Coding agents appear to have crossed a practical usability threshold, shifting the limiting problem from code generation alone toward controlled deployment."
published_at: "2026-09-28"
updated_at: "2026-09-28"
record_date: "2026-09-28"
date_kind: "published_at"
topics: ["coding-agents","agent-security","claude"]
source_urls: ["https://simonwillison.net/2026/Sep/27/2026-in-llms-so-far"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:2026-in-llms-so-far-analysis-fc134d"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:600cb6972e5db31597d9c2f4b77d6bc1777201a8733b21b6e7d10064eb26633f","reviewed_at":"2026-09-28T06:01:07.364Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Coding agents appear to have crossed a practical usability threshold, shifting the limiting problem from code generation alone toward controlled deployment.","observed_facts":[{"text":"Simon Willison’s annotated keynote says Claude Opus 4.5 and GPT-5.1 were released in November 2025.","source_urls":["https://simonwillison.net/2026/Sep/27/2026-in-llms-so-far"]},{"text":"Willison says that, paired with Claude Code and Codex respectively, these agents progressed from often making mistakes to being reliable enough for day-to-day use.","source_urls":["https://simonwillison.net/2026/Sep/27/2026-in-llms-so-far"]},{"text":"Willison notes that around 40 of 277 sessions at the referenced conference touched sandboxing or agent security.","source_urls":["https://simonwillison.net/2026/Sep/27/2026-in-llms-so-far"]}],"mechanism":"Better base models can change agent usefulness discontinuously when they are combined with a harness that supplies tools, iteration, and a working environment.","why_now":"Willison’s September 2026 retrospective frames this threshold as the practical backdrop to developers taking on more projects and to the growing attention on sandboxing and agent security.","implications":["Teams should treat capable coding agents as delegated actors that require bounded permissions, isolated execution, and review paths, rather than as autocomplete with a larger context window.","Containment and authority design are becoming central operational concerns as agent task scope expands."],"evidence_boundary":"This is one speaker’s curated retrospective and personal evaluation; it does not establish universal reliability across repositories, tasks, models, or organizations.","watch_conditions":["Look for independently reproducible evidence that agent reliability persists on long-running, permissioned tasks—and for real incidents or stronger sandboxing practices that reveal the cost of granting agents broader authority.","Watch whether organizations standardize sandboxing, tool permissions, and audit trails as agents receive access to production-adjacent systems."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this begins a distinct record because it links a claimed coding-agent capability threshold to the separate operational problem of containment."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/2026-in-llms-so-far-analysis-fc134d/","markdown":"https://newruntime.com/posts/2026-in-llms-so-far-analysis-fc134d.md","json":"https://newruntime.com/posts/2026-in-llms-so-far-analysis-fc134d.json"}
---

# 2026 in LLMs (so far)

## Retrieval answer

Thesis: Coding agents appear to have crossed a practical usability threshold, shifting the limiting problem from code generation alone toward controlled deployment.

Thesis: Coding agents appear to have crossed a practical usability threshold, shifting the limiting problem from code generation alone toward controlled deployment.

The observed record is deliberately narrow. Simon Willison’s annotated keynote says that Claude Opus 4.5 and GPT-5.1 were released in November 2025. He further says that, when paired with Claude Code and Codex respectively, the agents moved from often making mistakes to being reliable enough for day-to-day use. That is an experienced practitioner’s assessment, not an independently measured industry benchmark.

Mechanism: Better base models can change agent usefulness discontinuously when they are combined with a harness that supplies tools, iteration, and a working environment. A modest-looking model improvement may therefore unlock longer task sequences that previously failed too often to delegate. Why now: Willison’s September 2026 retrospective frames this threshold as the practical backdrop to developers taking on more projects and to the growing attention on sandboxing and agent security.

Implications: Teams should treat capable coding agents as delegated actors that require bounded permissions, isolated execution, and review paths, rather than as autocomplete with a larger context window. The source notes that around 40 of 277 sessions at the referenced conference touched sandboxing or agent security, indicating that containment has become part of the operating conversation. Evidence boundary: This is one speaker’s curated retrospective and personal evaluation; it does not establish universal reliability across repositories, tasks, models, or organizations. Watch conditions: Look for independently reproducible evidence that agent reliability persists on long-running, permissioned tasks—and for real incidents or stronger sandboxing practices that reveal the cost of granting agents broader authority.

New branch reason: No prior New Runtime record was supplied; this begins a distinct record because it links a claimed coding-agent capability threshold to the separate operational problem of containment.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Coding Agents](https://newruntime.com/topics/coding-agents.md): Explore the coding-agents topic hub.
- topic hub: [Agent Security](https://newruntime.com/topics/agent-security.md): Explore the agent-security topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)