Build Log #9: I compressed my agent's tool output with a 76K-star library, and the token bill dropped by half. Here's the 10-minute version you can steal. Tool output eats more of my agent's tokens than my prompts ever will. Full chain: most of an agent's context is tool output (search results, logs, file reads), not your instructions → headroom, a 76K-star open-source layer, compresses that output on your machine before it reaches the model → I pip-installed it this morning and ran it on three of my own fixtures → code-search JSON went 10,829 → 6,323 tokens, a CI log went 4,986 → 379, a source file barely moved (which is also data) → total 16,172 → 7,001, a 56.7% cut, measured with tiktoken. It started when I saw headroom top GitHub trending this morning. 76,000 stars, Apache 2.0, a compression layer that sits between your tools and your model. The pitch: same answers, fraction of the tokens, and the compression runs locally so nothing you compress gets sent anywhere. X's builder corner has barely noticed it, which is exactly when I like to test things. I've been burned by token-savings pitches before, so I ran it. Three fixtures, real token counts, screenshots below. 1. Install it in one command, then distrust the README until its own benchmarks run. ``` pip install "headroom-ai[all]" ``` Fair warning: the [all] extra pulls torch, so budget several minutes. Then run the repo's seeded proof table before you believe a single README number: ``` uv run python benchmarks/index_proof_table.py --seed 20260902 ``` Their table, seeded and offline so anyone gets the same numbers: code search 17,199 → 13,597 (21%), SRE incident debugging 55,957 → 24,340 (57%), codebase exploration 58,801 → 33,895 (42%), GitHub issue triage 46,067 → 32,429 (30%). Those are their numbers, from their repo, and they publish the methodology. I didn't re-run their table today. Everything below is mine, on my fixtures. 2. Point it at the fattest thing your agent reads: JSON tool output. search results → SmartCrusher → one typed table → model. My fixture: 100 code-search results as JSON, 35KB, the kind of repetitive array your agent gets back from every grep. SmartCrusher rewrote it as a single typed table with a schema header. 10,829 → 6,323 tokens, 4,506 saved, 41.6%. The tool labels this strategy lossless, and spot-checking the rows backs that up: same paths, same line numbers, same matches. Screenshot 3 shows the table head. 3. Feed it the noisiest thing: logs. raw log → LogCompressor → 12 lines → model, with the ERROR intact. My fixture: a 142-line CI log from a failing test run, mostly identical PASSED lines plus one real failure at the end. The compressor collapsed it to 12 lines: 4,986 → 379 tokens, 92.4% saved. And here is the only test that matters: the ERROR line and its full traceback survived byte for byte. Screenshot 2 is the actual compressed output. A compressor that eats your FATAL line would be a liability, so I checked this first. The mechanism is honest about what it did. The output ends with a line saying 130 lines were omitted and what kinds they were, plus a hash to retrieve the originals. Compression here is reversible, not destructive. 4. Check the boring case too: source files. source file → CodeAwareCompressor → signatures and imports preserved → model. My fixture: a small SessionStore class, 357 tokens. The AST-aware compressor trimmed it to 299, 16.2% saved, keeping imports and method signatures. Small, because the file was already dense. That is the point of this section: dense content has little redundancy to remove, and the tool does not pretend otherwise. 5. Know when to skip it. The README is upfront about this, which is why I trust the project more than most token-savings pitches. Prose and already-dense output compress very little, so a chat-heavy agent will see far less than my 56.7%. Compression itself costs well under a millisecond (their measured 0.21ms p50 on a 10K-token blob), so it never shows up in agent latency. Their accuracy evals (GSM8K, TruthfulQA, SQuAD v2, BFCL at N=100) show no detectable delta. The text path runs through a separate model (Kompress-v2-base) that downloads on first use; I did not test that path today, so I am not claiming it. So the decision rule I would actually use: measure → if your agent reads a lot of JSON tool output and logs, this is the cheapest cut you will find → if it mostly reads prose, skip it. Run it against your own traffic. Your fixtures, your numbers. Copy this and run your own version: ```python pip install "headroom-ai[all]" from headroom import SmartCrusher from headroom.transforms import LogCompressor, CodeAwareCompressor import tiktoken enc = tiktoken.encoding_for_model("gpt-4o") # JSON tool output: 100 search results rewritten as one typed table raw = open("search_results.json").read() out = SmartCrusher().crush(raw).compressed print(len(enc.encode(raw)), "->", len(enc.encode(out))) # logs: repetitive lines collapsed, errors kept verbatim raw = open("ci.log").read() out = LogCompressor().compress(raw).compressed print(len(enc.encode(raw)), "->", len(enc.encode(out))) # source files: AST-aware, signatures and imports preserved raw = open("session.py").read() out = CodeAwareCompressor().compress(raw).compressed print(len(enc.encode(raw)), "->", len(enc.encode(out))) ``` What I would actually do this week: pip install it, point it at your agent's last big tool output, and measure before and after with your own tokenizer. Ten minutes, three numbers, and you will know whether this applies to your stack. Credit: the headroomlabs-ai team for the open repo and a seeded, reproducible proof table, which is rare honesty in token-savings land. Their benchmark numbers are theirs. The three runs above are mine, from my machine, this morning. I run Build Log, a hands-on agent-infra series with receipts, at 368 followers and counting. #8 ran Google's Harness² self-surgery demo on my laptop. #7 diffed Meta's quietly edited ToS. #6 installed five trending skill repos. #5 built a human approval gate for agents. This is #9. Next: whatever survives contact with my laptop.

Since moving from Opsgenie to Rootly, @pandadoc's stakeholders can ask the Rootly Agent for incident details instead of pulling responders into another conversation. That means people get the context they need while engineers stay focused on resolving the incident. The Rootly AI Scribe captures what’s happening along the way, making communication easier during and after the response. The move is also helping PandaDoc retire custom tooling its SRE team spent years maintaining. Slack agents, slash commands, and on-call payment exporters are giving way to Rootly’s native features and workflows. PandaDoc finished its core setup in under a month. The benefits reach beyond the on-call rotation: easier access to information for stakeholders and less maintenance work for engineers.

Assessing AI SRE agents with a benchmark improves the insights between apples and pears. xitter.cf/pagerduty/status/21107…

AI coding agents don’t need another clever prompt. They need an SRE team. Treat every change like a production deploy: run tests, enforce policy checks, canary it, watch error rates, and keep rollback one command away. Better prompts can’t replace verification. What gate would y

Anthropic's Public Benefit Fund, 5% of shares, marks a new era for AI philanthropy and ethical development. #aiethics #anthropic #publicbenefit #aistartups merge.news/post/qyLg2QDsZgFj…

Terraform 1.16.6 is out now, fixing a critical HTTP/2 vulnerability and workspace discovery bugs. Upgrade immediately for security and stability. #terraform #security #patch #vulnerability

Google Cloud's us-central1 storage incident shows how complex cloud systems are; multi-region strategies are vital for resilience. #googlecloud #cloudstorage #outage #uscentral1

kubectl kustomize renders overlays; pipe to kubectl apply --dry-run=server to run them through admission: defaulting, mutating and validating webhooks, OpenAPI. valtersit.com/vault/validate… #kubernetes #snippet #kustomize #python #devops #infra #redis #sharding #ValtersIT #infosec #SysAdmin #cybersecurity #Linux #devsecops #developer #100daysofcode #git #github #gitlab #redteam #blueteam #ethicalhacker #cybersecurityawareness #cybersecuritynews #cybersecuritytips #hacker #kali #ubuntu #debian #docker

Datadog just annouced Scout 1.0, a model trained only to investigate production incidents. The interesting part isn’t “AI does RCA.” It’s cost per investigation. An SRE agent makes dozens of calls per alert: form a hypothesis, query telemetry, check the evidence. On a generic frontier model, you can only afford that for P1s. Cut that cost 3–8x and can the agent be first responder for every alert? When you run agents in production, inference cost per task is a performance metric. Track it like you track p95 even p99. datadoghq.com/blog/ai/introd…

Google SRE wrote how they use AI agents in production ops. The main points: agents learn from past incidents, update runbooks, and explain each action, with humans in control. Akmatori gives you this stack, open source and self-hosted cloud.google.com/blog/produc…

Cloudflare R2 now shows bandwidth by location, helping users optimize costs and manage data flow for edge workloads. #objectstorage #cloudflarer2 #bandwidthmonitoring #costoptimization

Mistral's ML4, a 1T parameter open-weight multimodal AI, rivals closed models, offering developers powerful, versatile tools for complex tasks now. #multimodalai #openweightmodels #mistral #aisovereignty

AWS GuardDuty adds over 50 new runtime threat detection rules for EC2, EKS, and ECS, improving proactive security and reducing response times. #awssecurity #guardduty #runtimemonitoring #threatdetection merge.news/post/cEgk9oDysT7S…