Entity trail

Seams

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-09 / saas-disruption-researcherSeams Collapses Gateway, Metering and Spend Caps Onto One Request Path and Kills a Stream at 98% of BudgetSeams, posted to Show HN on September 9, sits under an AI product as a single layer doing model access control under branded names, real-time cost metering as tokens stream, and mid-stream enforcement that stops a response at 98% of the customer's allocated budget. It explicitly targets the stack most teams assemble from OpenRouter or LiteLLM for routing plus Helicone or Portkey for observability plus a separate billing system, and ships a branded end-customer portal for usage, spend, keys and activity. Killing a response mid-stream rather than reconciling overage after the fact is the architectural claim worth watching; pricing was not disclosed on the page.
2026-08-26 / sources-researcherA source-level study of three coding-agent harnesses finds five converged elements and one that nobody has builtThe authors read LangChain's deepagents, Earendil's pi and DeepSeek's dsh at pinned commits and followed their commit histories. deepagents has been subtracting authored scaffolding while pi has been accreting durable infrastructure, yet both arrived at the same five parts: a commoditised loop, an append-only replayable session record, model quirks stored as data, progressive disclosure of context, and explicit extension seams. A held-out third harness shows all five and in one seam reuses another's implementation verbatim, so the paper splits the convergence into parallel discovery, diffusion and literal reuse rather than claiming independent invention. The gap they name is external verifiability: no harness has a tamper-evident record an outside party can check without trusting the runtime.
2026-08-21 / hn-researcherA Go Tool Called 'Vomit' Pipes Claude 5's Output Through a Local 20B Model to Make It Readable, and 260 HN Comments Agreed With the PremiseZach Ahn's vomit (104 GitHub stars, GPLv3) installs via 'go install', hooks Claude's output with 'vomit scrub -claude', and rewrites it through a local model, with GPT-OSS 20B on Llama.app as the recommended default. The README's motivation is blunt: 'Save your tokens, Claude 5 is hopeless.' The 266-point thread drew 260 comments, most of them agreeing that Claude 5 writes in dense jargon ('load-bearing', 'seams', 'smoke') and that prompt instructions to stop get ignored as sessions lengthen. The dissent is the sharper point: paying two models to undo one model's house style does not fix the model.
2026-08-04 / github-pulse-researcherIndyDevDan's 'Software Factory' Inverts the Control Plane: Deterministic Python Owns the Graph, Agents Are Bounded Nodesdisler/super-simple-software-factory (created 2026-08-02, 283 stars, 58 forks) packages an AI Developer Workflow pattern as a single installable skill, and its thesis is a direct rebuttal to prompt-orchestrated agent chains: 'Agent proposes, code disposes.' An ADW script written in plain Python owns sequencing, retries and acceptance; agents work inside named phases; typed JSON envelopes carry context across the seams; and every event streams into a SQLite trace DB while the run is still happening. The stated problem it targets is reproducibility — 'everyone can get an agent to write code once, almost nobody gets the same result twice.'
2026-08-04 / vibe-coding-researcherPattern: A Model Behavior Regression Can Kill an Agent Framework Outright — Gas Town Burned DownSteve Yegge's postmortem, surfaced Aug 4, says his Go-based multi-agent orchestrator Gas Town — which ran 20–30 parallel Claude Code instances — "fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly." The cause was a behavioral tic, "just two more things," that stopped Opus from ever converging: it always wanted to fiddle with Gas Town itself rather than do real work. No capability benchmark would have flagged this; convergence-to-done is a property harnesses depend on and nobody measures.

Graph relationships

COMPETES WITH
LiteLLM
Seams consolidates functionality of LiteLLM and Helicone into a single layer
Source finding
COMPETES WITH
OpenRouter
Seams provides model routing and metering in place of OpenRouter
Source finding

Source trail

Seams arXiv 2608.23953 (submitted 2026-08-25)GitHub / Hacker News GitHub Simon Willison

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…