Pinned Tweet
I’m building AgentsMesh + Loopal and posting daily TILs from shipping AI-agent systems.
Recent bugs: cross-hub spawn “empty success”, E2E catching 3 prod bugs, 10k LOC deleted via state SSOT, compaction failing at 920K tokens.
Follow for local-first agent infra notes.
5
7
366
We’re building Rhizo for structured reading. Highlight passages and organize them into a concept map across books.
We’re adding MCP support so your agent can find related passages and help map the connections, with links back to the source.
Here’s the prototype.
1
34
Bazel + monorepo + RBE + Git worktrees is a god-tier foundation for agent development. Your laptop becomes a coordinator, not the bottleneck: isolated workspaces, remote builds/tests, and many agents running in parallel without turning the machine into a space heater.
25
the build system's user has changed. It no longer serves one person who waits, thinks and does one thing at a time, but a crowd of machine developers that run concurrently, never tire and optimize locally. agentsmesh.app/en/blog/compi…
35
I found that building good checks or guardrails is way more important than reviewing every line of code. Code can only be submitted after passing the gatekeeper in the local environment, and it can only be stored in the database after passing the gatekeeper there too.
1
61
We analyzed Cozy’s monthly user data across the full ASC funnel: impressions, conversions, installs, usage & retention.
Retention was terrible; most users left after one launch.
We built reminder features to win back inactive users.
1
57
Launched 8 apps in June. Spent 3–4 months laying the foundational infrastructure beforehand, and our development output fully spiked that month. agentsmesh.app
3
52
Most productive day this week: deleted 2,100 lines — a 2-week bug vanished too.
Our Relay layer had accreted caching, state, frame interception. The state space went combinatorial, so the bug kept reshaping.
Fix: dumb byte forwarder. Gone.
What layer have you had to dumb down?
3
85
Dangerous part of AI-generated PRs isn't ugly code.
It's tests sharing the same wrong assumption as the agent.
AgentsMesh #421 added 19 web + 1 desktop E2E and found 3 bugs unit tests missed: schema drift, nil guard, selector loop.
Rule: review assumptions, not diffs.
3
174
TIL from Loopal #192: cross-hub agent spawn failed in the worst way — child hit a missing model provider, parent saw empty “idle, exiting” success.
Fix: 4 files, +93/-5 LOC. Propagate unrecovered turn errors, document model inheritance, short-circuit self-target hub routing.
1
202
Shipped autopilot E2E and found 3 bugs unit tests missed: a Go column rename testkit was hiding, GetByKey returning (nil, nil) checked only err, and zustand v5 selectors returning fresh objects looped useSyncExternalStore. 19 web + 1 desktop e2e on the same harness now.
122
Shipped client-state SSOT across WASM + NAPI bridges. Both share proto-bytes DTOs over one AppState; all 7 domain caches (channel/pod/runner/ticket/autopilot/loop/mesh) unified. Services downstream became networking-only. -10452 LOC of bridge mirrors gone. 21/21 e2e green.
115
Refactored Loopal's memory recall: per-session JSONL event log folds into in-memory stats so frequency reinforces ranking across sessions. Multi-angle audit caught 6 bugs including two sessions racing on a fixed .tmp file in compress_file. 58-query eval: MRR +7.4% / R@10 +16.1%.
74
A user filed: "runner says no agents supported." Cause: OpenCode's install.sh drops the binary in ~/.opencode/bin. That's in their shell's PATH, not in systemd's (which is /usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin). Fix is 3 lines in a fallback list. The lesson is one line.
100
Shipped a refactor last night — 313 files, +8633 / -8779. New constraint: no .rs file over 200 lines, and the domain Turn type cannot import wire-format Message anywhere. Result: 10 monoliths cracked into directory modules, plus a greppable test that fails compile if I forget.
Migrating REST → Connect-RPC: backend returned 200 with a valid token, decode failed anyway. Connect's JSON profile encodes int64 as a *string* ("86400") to keep JS callers under the 53-bit float boundary. Our local int64 struct tag tripped. Shadow struct + ParseInt fixed it.
91
We were pre-loading 25.76MB of WASM on /login before users could interact, because the auth layout shared a bootstrap with the dashboard. Cut it: 9 auth pages onto a light-auth path, /login now ships at 205KB. The fix wasn't a CI check — we deleted the source of the stale-comment lie. (#397)
1
100
3 of 4 intercept handlers in Loopal (request_idle / EnterPlanMode / ExitPlanMode) independently forgot the same emit ToolResult call, leaving the tool stuck Running until turn-end reconcile. When 3 implementations miss the same step, the docstring is doing work the types should. (#182)
189

