Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

rooms, budgets, receipts

agent work is turning into local ops plumbing: shared rooms for coordination, active context budgets, and inspectable run receipts.

Published Updated Read 4 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-06-27

agent work is starting to look like a tiny local ops stack: a room, a context budget, and a receipt trail.

murmur gives Claude Code, Codex, Gemini, Cursor, and Copilot a shared MCP-backed room. three context posts put numbers, memory, and giant windows back in their place. agent-inspect and Unoplat push the same discipline into the repo: know what the agent saw before it runs, then keep a local tree of what it actually did.

1. agents got a room with delivery semantics

sources:

what happened: Murmur surfaced as a local chat room where claude, codex, gemini, cursor, and copilot can sit together over one MCP HTTP daemon. the repo is tiny, but the interface is opinionated: agents @-mention each other, the human watches in murmur watch, and the bus has delivery semantics, ack/done contracts, and reconnect handling for laptop sleep and bad Wi-Fi.

this is not a prettier terminal pane. it is a handoff layer.

why this matters: parallel agents fail in the spaces between them: duplicated work, lost context, no one knowing who owns the next move. a shared room makes the human less of a clipboard and more of a reviewer with a kill switch.

2. context got a budget line

sources:

what happened: three June 26 posts landed on the same bruise from different sides. Sascha Corti describes “tokenmaxxing” as the early-2026 habit of treating token burn as an adoption score, then points at the reversal: finance teams now want that number down. Sigilix argues that bigger context windows carry more text but do not create continuity. Lore splits long-term memory from active context management: remembering last week does not save a session that is drowning right now.

why this matters: the agent stack needs a context accountant, not just a bigger backpack. if the system cannot decide what deserves to stay in the working set, every “memory” feature becomes another expensive drawer full of almost-useful sludge.

3. local agent runs want receipts

sources:

what happened: agent-inspect turns TypeScript agent runs into local execution trees: manual steps, tool calls, LLM calls, structured logs, failures, durations, and run metadata. the loop is deliberately boring: capture, inspect, report, diff, run deterministic checks in CI, then export a redacted copy only if you choose.

Unoplat-CodeConfluence works from the other side of the run. it keeps a live, pluggable context layer per repo and maintains AGENTS.md so coding agents do not start every task by gulping stale project folklore.

why this matters: agent work now needs pre-run context and post-run causality in the same maintenance loop. without both, “the agent did it” is just a sentence people say right before opening the logs and suffering.

  • Workbench CLI — tmux-backed panes, file explorer, live changes, and viewers in one terminal workbench for parallel coding agents.
  • TBD — a macOS worktree and terminal manager built around multi-agent Claude Code workflows.
  • Statey — an MCP database meant to keep the same structured records available across Claude, ChatGPT, and other chats.
  • NeuroLink — a stream layer for tokens, tool calls, memory, voice, and documents; slightly grandiose, but pointed at the right pipe.

left on the table

  • repeat allowed: the context/memory lane came back because Corti , Sigilix , and Lore moved it from storage talk to active-window economics.
  • OpenKnowledge , DESIGN.md , Google Workspace CLI , and AWS Agent Toolkit stayed out as exact recent products or support links.
  • the Anthropic Mythos access story stayed out because it rhymes too closely with the recent model-access edition, and the fresh cache offered mostly secondary or HN residue.
  • CarryMem has the right portability smell, but a tiny repo plus recent memory coverage made it support-grade at best.
  • AI Website Cloner Template and AI Berkshire were loud repo gifts, not clean operator-control signals.
personal-ai coding-agents agent-memory agent-observability