01
Edge runtime
Workers on a global edge network. Cold start < 5ms, instant rollback via atomic deploys, no origin auth round-trip.
- Cloudflare
- WASM
- KV
// engineering lab · v3.2 · uptime 99.97%
Build logs from a quiet lab.
t3ch.b3rt.dev is the engineering arm of the empire — system design notes, reproducible experiments, and the day-by-day build log of a one-person R&D shop. Plain text, plain graphs, no marketing.
214Commits this quarter
17Active services
6.2 msP95 latency
0Incidents / 90d
Recent commits
- a3f9c1e refactor(cron): collapse four queues into one priority lane
- 7e1d4b2 feat(agent): add subagent launcher with bounded context
- c402a88 perf(sites): cut empire hub cold start to 41ms via edge cache
- f1b3d09 fix(voice): prevent cua-driver foreground steal on retry
- 9c8a7e5 docs: publish deployment runbook for empire stack
- 4b9e2c1 feat(brand): ship quarterly issue No. 04 art.b3rt.dev
- 2d6a0f4 chore(deps): bump hermes-agent to 0.14.2, security patch
- 8e7b351 feat(skills): add hermes-agent-skill-authoring playbook
- 5c1f7a0 fix(llm): cap streaming tokens to prevent OOM on long tasks
- 1a4d29c feat(monitor): wire empire heartbeat to pager on 5xx spike
- 6f0e8d3 test(agents): add replay harness for queued agent runs
The stack
Boring where it should be, novel where it earns its place.
02
Agent mesh
A bounded set of named agents on a single priority queue. Cron-driven, human-gated, with replayable job ledgers.
- Hermes
- Cron
- SQLite ledger
03
Static sites
Hand-rolled HTML + CSS, zero JS frameworks. Sub-15KB CSS, sub-25KB HTML, served from edge cache. Fast because there is nothing to load.
- HTML5
- CSS only
- SVG art
04
Observability
Structured logs, a heartbeat cron, and one public status page. If a metric moves, a human is paged within five minutes.
- Pings
- Logs
- Alerts
05
LLM routing
A tiered model router. Cheap structured work goes to a quantized local GGUF on the AMD box. Hard reasoning routes to a frontier model. Latency-budgeted.
- GGUF
- Router
- Budgets
06
Brand tokens
One CSS file of design tokens drives every site. Swap a single custom property to retune the empire. No per-site build, no compile step.
- Tokens
- No build
- CSS vars
Current experiments
Things that might become features, or might get deleted.
Subagent task ladder
Decompose a large build into bounded, claimable units. Each unit ships with a contract, a verifier, and a rollback plan. Staged rollout across three sites this quarter.
Browser with intent
A typed browser action surface for agents. Clicks are addressable by element ref, not pixel coordinates. Failures escalate through a verify-rather-than-retry ladder.
Local LLM residency
Routing cheap, structured work to a quantized local GGUF on the AMD box. Goal: drop external-API cost per empire task by 40% without raising latency.
Replay ledger
Every agent run gets a deterministic replay from its ledger entry. If a run misbehaves, we replay it locally, find the diff, and patch the agent rather than the symptom.
Edge KV namespace per site
Move per-site counters and feature flags from origin to edge KV. Sub-millisecond reads, zero origin load, atomic rollouts via versioned keys.
Operating principles
Five rules that govern every change to the empire stack.
-
i.
Boring wins.
Prefer the tool you have already mastered. Novelty is a cost; it must earn its place against the maintenance tax.
-
ii.
Small surface, deep seams.
Few endpoints, well-typed. The interface is the product; the implementation can be rewritten.
-
iii.
Humans gate merges.
Agents propose, humans dispose. Every shipped card has a human reviewer named in the ledger entry.
-
iv.
Trace everything.
If you cannot replay a run, you cannot debug it. Logs are append-only. Decisions cite their evidence.
-
v.
Refuse the seventh.
One type family, two weights, one accent. The seventh idea is the one you say no to.