Leading AI practices. Agentic network engineering https://teapawt.net/t.co/j0Gu0Qq4RZ Systems not the prompts https://teapawt.net/t.co/a2Ram1Rarv
MD
Joined November 2015
- Tweets7.4K
- Following791
- Followers1.4K
- Likes3.4K
5. Minimizing usage: structure, not willpower.
A knowledge bank: findings get IDs, and agents cite IDs instead of re-researching.
Narrow briefs: a sub-agent gets the files to read, the goal and what’s ruled out. It returns a 15-line summary, not file dumps.
Small batches (at most 10 things for human review) and automated gates before any human time.
Hooks warn on untyped agents and heavy jobs started by day, and block writes to protected data.
Measure it: log tokens per model per session, and re-check spend on a date.
4. When I use Fable: rarely, and only by name.
When I'm stuck on a problem that i can solve i use fable and break it down into as close a principle i can and use that to build on.
3. Model selection: the cheapest tier that can judge.
Mechanical (run, count, move, queue, read a summary): the small model (Haiku).
Judgment against sources (research, analysis, writing, review of data): the mid model (Sonnet).
Design, hard code, adversarial and security review: the top model (Opus).
Minimize sub agents as much as possible and dont use general agents as they default to the parent so you could be using opus for task that needs hakiu
2. What goes on the queue: anything long, heavy or repeatable.
If it runs over a few minutes, hogs a resource (GPU, CPU, network, a drive), or repeats nightly, it’s a queued job: a plain command with a resource class, a resume key and a time window. It isn’t an agent.
Prove one item in the foreground (--limit 1, --dry-run); queue the batch. Agents decide what to run; code runs it. Never pay a model to babysit a loop.
It would take me about 30 min to explain, but short answer is.
Workflow: one conductor, one queue.
One main session owns the plan and a live work queue. Each row has an owner, a status, and “superseded by”.
Before starting anything, check whether newer work made it moot. Most wasted spend is work that was already dead. Rabbit holes get parked as “deep dive later”, not chased. Finish first, iterate later.
One day I'm an audio engineer working compression, next working on AWS automating deployments with Terraform, then working with security later that day on fraud and billing tools. Then running analytics for a call center with ML to help find problematic customers.
100 percent. Opus 5.5 had major changes in how it works. It checks its own work and runs its own loops. So the need to use subagents has reduced but not completely gone away.
Single agent loops will be the thing in 2027 in frontier models. If you are still doing multi agent for new projects you are behind the curve
Once you learn how to really use claude you truly have a superpower if you thought what i was doing was crazy last year wait…..
These are good, but you have to maintain struct enforcement of the sub task running or they will drain your tokens pretty fast if you aren't paying attention to what they are there for.
Boris Cherny, Head of Claude Code at Anthropic, just shared an official tip before Fable 5.5 ships
The setup: split your workload across Opus 5.5, Sonnet 5.5, and Haiku 5.5, and let them talk directly across your open terminals
Zero context pollution between windows: independent sessions ping each other via plain-text IPC over local sockets without burning a single token on conversation history
You just launch Claude Code like this:
claude --model sonnet --settings '{"crossSessionInbound":"accept"}'
here's the split:
→ Opus 5.5 architects session permissions: audits ListAgents discovery, enforces inbound message policies, and signs off on multi-terminal diffs
→ Sonnet 5.5 executes code fixes: builds endpoints, updates schemas, and patches dependent code whenever a peer session reports a breaking change
→ Haiku 5.5 relays inter-session messages: processes SendMessage pings across terminals at 340 tok/s with zero history bloat and sub-millisecond local socket latency
→ you stop burning Opus context on manual copy-pasting between terminals and keep your sessions lean
Opus only steps in at three moments:
→ setting up cross-session topology and inbound security rules before launch
→ evaluating contract diffs when a peer message signals an API schema update
→ signing off on synchronized multi-session PRs before merging to main
Opus architects · Haiku relays · Sonnet executes
JEV engineering is the same move one layer down: the forks that need no thinker (which peer, which tool, forward or drop) go to JEV in milliseconds, and the big models only see the ones that split
the full tree
> Opus 5.5 architects session permissions and audits cross-session topology
> Haiku 5.5 relays SendMessage pings between terminals at 340 tok/s
> Sonnet 5.5 patches code across workspaces when a peer signals a change
> Sessions discover each other automatically via ListAgents and /peers
> Messages travel as isolated plain text over local IPC sockets with 0 context dump
> JEV routes micro-forks in milliseconds
paste the tree and this prompt into Claude Code ↓
"Rebuild my Claude Code setup around this tree:
1. Configure cross-session discovery using /peers and set crossSessionInbound to accept
2. Delegate all peer-to-peer message routing and notification triage to Haiku 5.5 at 340 tok/s
3. Spawn Sonnet 5.5 to execute code edits and synchronize types whenever a peer session reports an update
4. Restrict all cross-session payload traffic to lightweight plain text via SendMessage with zero history transfer
5. Reserve Opus 5.5 for inter-session capability authorization and final sign-off before committing
Show me every change as a diff first. No edits until I say go"
↳ code.claude.com/docs/en/cros…
bookmark this, your terminals now talk to each other
Donte retweeted
Decided to create a new one. youtube.com/@deirradecoded
I use terraform/ansible a lot to build in aws. Terraform is great for buildouts ansible is like my day 2 tool
