Mark Ericksen

Mark Ericksen

@brainlid.bsky.social

Keeping an open mind, sharing what's on my mind. ThinkingElixir.com, podcast.ThinkingElixir.com, #ElixirLang

1.3K followers129 following124 posts

Mark Ericksen
@brainlid.bsky.social

More polish: - StreamingSession: host-agnostic streaming helpers with multi-tool-safe deltas, so sibling tool calls keep their UI state - Middleware can persist synthetic transcript entries (like multi-step processes) through the same pipeline as LLM messages

Mark Ericksen
@brainlid.bsky.social

v0.8.0 has breaking changes: transport, SubAgent task_name, Session/Factory, integer todo ids, replace_file_lines removed. Easiest upgrade: re-run mix sagents.setup on a clean branch and merge changes in. Full guide in the changelog: github.com/sagents-ai/sagents

github.comGitHub - sagents-ai/sagents: Build interactive AI agents in Elixir with OTP supervision, middleware composition, human-in-the-loop approvals, sub-agent delegation, and real-time Phoenix LiveView integ...Build interactive AI agents in Elixir with OTP supervision, middleware composition, human-in-the-loop approvals, sub-agent delegation, and real-time Phoenix LiveView integration. Built on LangChain...
Mark Ericksen
@brainlid.bsky.social

The generated Coordinator had grown to ~390 lines of session boilerplate living in YOUR app. v0.8 moves it into the library: Session, Factory, FactoryRouter. Per-request data flows through a typed config. Your Coordinator drops to a thin bridge.

Mark Ericksen
@brainlid.bsky.social

Quality of life: - Pin allow_other / allow_cancel on ask_user instead of trusting the LLM each call - Todo ids are integers now, so "10" stops sorting before "2" - Optional inline todo snapshots rendered right in the chat transcript

Mark Ericksen
@brainlid.bsky.social

An agent crosses 3 process boundaries per run. OpenTelemetry spans, Sentry context, tenant scope, Logger metadata: all lost across them. Detached traces, nil tenant lookups inside tools. New ProcessContext middleware carries it all the way to your tools.

Mark Ericksen
@brainlid.bsky.social

v0.7 sent agent events over Phoenix.PubSub: a cluster broadcast bus for what's usually ONE local subscriber. Streaming deltas scaled as nodes x agents x tokens/sec of cross-node traffic. v0.8 uses direct, monitored delivery. Same API, far less noise.

Mark Ericksen
@brainlid.bsky.social

Structured output used to mean bypassing your agent: losing token tracking, scope, tenancy, fallback models, all your middleware. New Sagents.Extract runs extraction THROUGH the agent. Define a schema, get typed args back. Full context, no AgentServer needed.

Mark Ericksen
@brainlid.bsky.social

Extraction or tool output failing validation? New :until_tool_success keeps the run going on a bad result so the LLM can fix its own arguments, bounded by max_runs. Errors feed back, the model retries, you get a clean result. Validate-and-retry, built in.

Mark Ericksen
@brainlid.bsky.social

Your agent asks the user a question, then the server shuts down: deploy, idle timeout, crash. v0.7: the question was silently lost on reload, conversation came back as if nothing was pending. v0.8: the agent boots back into :interrupted with the question intact.

Mark Ericksen
@brainlid.bsky.social

Need a tool that STOPS a run, not pauses it? A validation gate that says "not ready, stop here"? LLMs ignore polite stop requests and plow ahead trying to help. New :halt interrupt via the Haltable middleware is a hard terminal stop. The model doesn't get to argue.

Mark Ericksen reposted

Lars Wikman
@lawik.bsky.social

Big congrats and props to Thinking |> Elixir for 300 issues of continuously and consistently useful podcasting. podcast.thinkingelixir.com/300 Showing up like this is a skill and they've done a great job at it. #elixirlang

podcast.thinkingelixir.comCelebrating 300 with José ValimEpisode 300! News includes the Localize library v0.15.0, a new program dependence graph tool called Reach, Tidewave voice input and PWA support, ERLEF 2026 elections, plus a special in-depth…
Mark Ericksen
@brainlid.bsky.social

The agents_demo app has been updated to work with Sagents v0.4.0. Full-featured chat UI with real-time agent execution, file browsing, and conversation persistence -- all running on the latest release. #ElixirLang #MyElixirStatus

Mark Ericksen
@brainlid.bsky.social

Metadata-only changes (tags, title) no longer rewrite the full file to storage. A new dirty_non_content flag routes to update_metadata_in_storage when the backend supports it, falling back to full write transparently.