2026-04-24

• #0010

  #0010Models & APISimon Willison4 weeks ago

  DeepSeek V4 previews push frontier-level open weights at unusually low cost

  100radar

  DeepSeekAI model family — open weights and low cost

  DeepSeek released two preview models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, with 1M-token context, open weights, and MIT licensing. From an indie developer perspective, the main signal is the price-performance tradeoff: near-frontier capability appears to be getting dramatically cheaper, even if local deployment remains difficult due to very large model sizes.

  • —This is a new model release, so its importance is high for developers tracking model options.
  • —V4-Pro and V4-Flash combine very large MoE architectures with open weights and an MIT license.
  • —The biggest takeaway for indie developers is lower inference cost, which can expand experimentation and product margins.
  • —Practical local use is still constrained because the model files are extremely large, especially for Pro.

  Source: simonwillison.net/2026/Apr/24/deepseek-v4/#atom-everythiRead original →
• #0009

  #0009Models & APISimon Willison4 weeks ago

  GPT-5.5 launches in Codex and ChatGPT, while API access remains pending

  100radar

  CodexCoding agent service — ChatGPT-linked path for experiments

  The post frames GPT-5.5 as a strong new model already available in Codex and rolling out to paid ChatGPT users, but notably not yet in the public API. For indie developers, the key takeaway is that access may currently depend more on subscription-linked Codex paths than on standard API workflows, which creates both opportunity and platform risk.

  • —GPT-5.5 is presented as a fast and highly capable new model, which makes this notable as a model-release signal.
  • —The standard API is not available yet, so developers cannot rely on the usual clean API-based evaluation and product integration path.
  • —A semi-official Codex-specific endpoint tied to ChatGPT subscriptions appears to be tolerated or supported, which may lower short-term experimentation costs for indie hackers.
  • —That access path is strategically useful but fragile, because it depends on product-specific mechanisms rather than a fully announced public API contract.

  Source: simonwillison.net/2026/Apr/23/gpt-5-5/#atom-everythingRead original →
• #0008

  #0008Agents & toolsCursor Changelog4 weeks ago

  `Cursor` adds async subagents, better worktrees, and multi-root workspaces

  80radar

  Parallel agent work, cleaner worktrees, and cross-repo work in one workspace reduce context-switching fast. Useful now if you split products across app, backend, and infra repos.

  • —Async subagents add parallel task handling inside the editor, which should speed up research, edits, and review loops.
  • —Improved worktrees make branch-isolated experiments less painful, so risky changes get easier to test without polluting the main tree.
  • —Multi-root workspaces let you edit across repos in one window, a strong fit for app plus API plus infra setups.
  • —The bundle targets real workflow friction rather than model novelty; teams are not required to benefit from it.

  Source: cursor.com/changelog/04-24-26Read original →
• #0007

  #0007Agents & toolsHermes Agent Releases4 weeks ago

  `Hermes Agent` `v0.11.0` turns its CLI into a real interface layer

  80radar

  Hermes AgentAgent CLI platform — TUI, multimodel, plugin-native

  The CLI is now a React/Ink TUI on top of a JSON-RPC backend, while transports, plugins, and provider routes were widened across the stack. This is a substantial usability and extension release; worth testing now if you want one agent shell to span local, cloud, and chat surfaces.

  • —hermes --tui was fully rebuilt with React/Ink plus tui_gateway, adding live streaming, sticky composer, OSC-52 clipboard, status bar, and subagent visibility. The terminal UX moved from wrapper to primary workspace.
  • —A new agent/transports/ abstraction split provider formatting from run_agent.py, and native AWS Bedrock landed on top via Converse API. Multi-provider support got structurally cleaner, not just broader.
  • —Five inference paths were added at once: NVIDIA NIM, Arcee AI, Step Plan, Gemini CLI OAuth, and Vercel ai-gateway; Gemini also gained native AI Studio routing. Provider optionality jumped fast.
  • —GPT-5.5 is exposed through Codex OAuth with live model discovery in the picker, so new OpenAI models can appear without manual catalog updates. That cuts maintenance friction.
  • —Plugins now register slash commands, dispatch tools, veto tool calls, rewrite tool results, transform terminal output, ship image_gen backends, and add dashboard tabs. The plugin layer crossed from hooks into platform territory.

  Source: github.com/NousResearch/hermes-agent/releases/tag/v2026.Read original →
• #0006

  #0006Agents & toolsClaude Code Releases4 weeks ago

  `Claude Code` `v2.1.119` improves config persistence and enterprise PR flows

  60radar

  Claude CodeCoding agent CLI — stronger PR and policy consistency

  Config now persists across ~/.claude/settings.json, agent permissions align between print and interactive modes, and PR ingestion expands beyond GitHub. Daily CLI friction drops enough to justify updating now.

  • —/config settings now persist in ~/.claude/settings.json with project/local/policy precedence, so shared defaults stop fighting repo-specific overrides.
  • —--from-pr now accepts GitLab MR, Bitbucket PR, and GitHub Enterprise URLs, while prUrlTemplate lets footer badges point at internal review systems.
  • —--print now respects tools: and disallowedTools:, and --agent <name> honors permissionMode; non-interactive runs behave closer to real agent policy.
  • —MCP server reconfiguration connects in parallel, plugin version pins auto-update to the highest compatible tag, and PowerShell auto-approval reaches parity with Bash.
  • —The rest is polish with real payoff: paste/newline bugs, fullscreen scroll snaps, OAuth discovery failures, missing Glob/Grep, and stale spinner states are fixed.

  Source: github.com/anthropics/claude-code/releases/tag/v2.1.119Read original →
• #0005

  #0005Agents & toolsHacker News · Show HN AI4 weeks ago

  `SuperHQ`: Run Coding Agents in `microVM` Sandboxes

  60radar

  SuperHQAgent platform — runs coding agents in isolated microVMs

  Each agent gets a full Debian microVM, mounts your project, and writes through a tmpfs overlay so the host stays untouched. The diff-accept flow and out-of-sandbox API key handling make local agent runs much safer; worth testing now.

  • —Every agent runs in its own full Debian VM, not a shared host process, so filesystem and process isolation are much stronger.
  • —Project folders mount into the sandbox, but writes land on a `tmpfs` overlay first, letting you inspect changes before they touch the host.
  • —Secrets stay outside the VM: API keys never enter the sandbox, which cuts one of the bigger risks in autonomous coding setups.
  • —remote.superhq.ai adds remote control for workspaces and agents, so the local sandbox model can extend to anywhere access without changing workflow.

  Source: github.com/superhq-ai/superhqRead original →
• #0004

  #0004OtherSimon Willison4 weeks ago

  Claude Code quality issues were caused by harness bugs, not model regressions

  50radar

  Claude CodeCoding agent — terminal-first AI dev assistant

  This post highlights that recent Claude Code quality complaints were traced to three harness-level bugs rather than weaker underlying models. For indie developers building agentic tools, the main lesson is that session management, context retention, and runtime orchestration can degrade product quality as much as the model itself.

  • —Anthropic's postmortem says three separate Claude Code harness bugs materially harmed output quality.
  • —A bug introduced on March 26 repeatedly cleared older reasoning in long-idle sessions, making Claude appear forgetful and repetitive.
  • —For indie developers, this is a reminder to monitor harness behavior, not just model quality, when debugging agent failures.

  Source: simonwillison.net/2026/Apr/24/recent-claude-code-qualityRead original →
• #0003

  #0003Models & APIAnthropic Python SDK4 weeks ago

  `anthropic-sdk-python` `v0.97.0`: `CMA Memory` Public Beta

  50radar

  CMA MemoryMemory API — manages conversation and user context

  CMA Memory landed in public beta, with multipart upload copying also optimized. Worth a look if you are building memory-heavy agents; no model or pricing change, so urgency is moderate.

  • —CMA Memory is now in public beta at the API layer, opening the door to stateful agent patterns without custom memory plumbing.
  • —Multipart request handling now copies file structures more efficiently, which should help large upload paths feel less wasteful.
  • —This release is mostly API polish: spec fixes, restored missing features, and bootstrap/test maintenance rather than headline platform changes.

  Source: github.com/anthropics/anthropic-sdk-python/releases/tag/Read original →
• #0002

  #0002Models & APIAnthropic TypeScript SDK4 weeks ago

  `Anthropic TypeScript SDK` `0.91.0` adds `CMA Memory` public beta

  50radar

  CMA MemoryMemory service — persistent state for agents

  CMA Memory is now exposed in the TypeScript SDK, with Bedrock auth handling also updated. Useful if you're building longer-lived agent state; otherwise this is a minor maintenance release.

  • —CMA Memory moved into public beta in the SDK, opening a direct path to persistent memory features from TypeScript apps.
  • —Amazon Bedrock integration now uses an auth header for the mantle client, which matters if you route Anthropic usage through Bedrock.
  • —Two API fixes landed together: spec errors were corrected and missing features were restored, reducing SDK drift risk.
  • —The rest is maintenance work: a more robust bootstrap script and test dependency bump to steady v0.22.1.

  Source: github.com/anthropics/anthropic-sdk-typescript/releases/Read original →
• #0001

  #0001Agents & toolsGitHub Changelog4 weeks ago

  `GitHub Copilot Chat` gets deeper PR and diff context

  40radar

  GitHub Copilot ChatAI coding assistant — deeper PR and diff context

  Chat now pulls in richer diff and pull request context while you review changes. If you already use Copilot, the review-to-fix loop gets tighter.

  • —Copilot Chat now works with richer context around diffs and pull requests, so answers should be less detached from the actual patch.
  • —The update adds new capabilities during PR work, signaling a shift from generic chat toward review-aware coding assistance.
  • —Access starts from asking questions on a pull request flow, which lowers friction versus moving context into a separate tool.

  Source: github.blog/changelog/2026-04-23-copilot-chat-improvemenRead original →