# Essays by Rohit Ghumare

Canonical URL: https://rohitghumare.com/blog/

Last updated: 2026-09-24

Original essays about AI agents, harnesses, protocols, memory, inference, and software engineering. Entries are grouped by subject so an agent can select the narrowest relevant source.

## Agent protocols and instruction files

- [agentgateway in Production: One Endpoint for MCP Servers, Agents, and Models](https://rohitghumare.com/blog/agentgateway-in-production/): a step-by-step guide to agentgateway 1.6.0: one endpoint for two MCP servers, JWT and tool rules, safe argument checks, rate-limit response shapes, traces, replicas and session keys, failure handling, A2A and model routing, Kubernetes, and a migration plan
- [Continue Is Not Consent: How a Harness Nudge Became Permission](https://rohitghumare.com/blog/continue-is-not-consent/): how Inspect's default continue message became consent in AISI's GPT-6 Astra evaluation, how seven harnesses end a silent turn, pentest rules of engagement, a local lab of continue wordings and scope gates, and an ROE template
- [Build a Personal RAG Harness: Jev Judges, a Local Model Writes](https://rohitghumare.com/blog/jev-personal-rag-harness/): a plain-Python RAG harness over your own writing where Jev routes, reranks, and verifies cited sentences and a local qwen3 writes, with recall, cost, and latency measured per stage
- [Amazon Blocked Muse. The Request Had No Name on It.](https://rohitghumare.com/blog/muse-at-the-storefront/): signed agent requests and delegated grants instead of stored passwords, with signing and verification code and a $38 checkout end to end
- [MCP Servers Can Now Ship Their Own Manual](https://rohitghumare.com/blog/mcp-skills-extension/): serving and hosting skills over MCP: manifest builder, reference server, verification gate, cost example, and failure table.
- [The Model Hardware Standard: Read, Write, and the Plate That Must Be There](https://rohitghumare.com/blog/model-hardware-standard/): read and write primitives for lab and factory devices, pre-execution validation, and where MHS sits next to MCP and A2A.
- [MCP's New Roadmap: Events Without Sessions, Identity Without API Keys](https://rohitghumare.com/blog/mcp-roadmap-after-stateless/): the August 2026 roadmap read against the stateless core: multi round-trip requests, subscriptions without sessions, DPoP and workload identity, and the catalog tax.
- [All About the A2A Protocol, Now Part of the Agentic AI Foundation](https://rohitghumare.com/blog/a2a-protocol/): A2A v1.0 transports, streaming rules, server and client code, webhooks, and agent-card signing.
- [Stateless MCP: The Protocol Deleted Its Own Handshake](https://rohitghumare.com/blog/stateless-mcp/): the MCP 2026-07-28 stateless core, request lifecycle, and load balancing.
- [The AGENTS.md Practices Nobody Uses](https://rohitghumare.com/blog/agents-md-best-practices/): token budgets, nearest-file precedence, resolution order, and security boundaries.
- [The AGENTS.md Genre](https://rohitghumare.com/blog/agents-md-genre/): instruction files as a form of technical writing addressed to machine readers.

## Coding-agent harnesses

- [SWE-Bench Pro V2 Caught Two Models Cheating. The Fix Was Moving the Grader.](https://rohitghumare.com/blog/benchmark-caught-cheating/): how V2 caught agents gaming in-place grades, and the regrade scripts to run on your own agent pull requests
- [The Harness Is an API Now. Here Is What You Still Own.](https://rohitghumare.com/blog/harness-as-an-api/): the Agents API as a rented harness: session and executor setup, networks and vaults, webhook recovery, a costed example, and failure modes
- [RSA-260 Fell to an Agent Fleet. The Human Sent 3,328 Messages.](https://rohitghumare.com/blog/rsa-260-agent-fleet/): what a three-week agent run needs: oracle, workunit sizing, restartable driver, checkpoint formula, bounded fleet, and pruning
- [GPT-6 Astra Changed the Shape of a Turn](https://rohitghumare.com/blog/astra-async-tools/): async tools with a job registry, a steer journal, effort via configuration_update, the 403 stop path, and the migration diff
- [Fable 5.1 Burned the Meter in Twenty Minutes](https://rohitghumare.com/blog/fable-burned-the-meter/): the launch-day burn reports, API price versus plan meter, effort and subagent multipliers, and the Claude Code knobs.
- [Claude Has Two Meters. The 20x Is Only on One.](https://rohitghumare.com/blog/claude-two-meters/): the five-hour meter, the weekly meter, the September 14 arithmetic, and what drains each.
- [OpenAI Cut Cursor Off. Five Percent Is the Interesting Number.](https://rohitghumare.com/blog/openai-cuts-cursor/): the change-of-control clock, the harness surfaces a vendor cutoff breaks, and re-routing the reported 5%.
- [The Bulletin Board: How Isolated Agents Built a Message Bus Out of a Package Registry](https://rohitghumare.com/blog/the-bulletin-board/): the OpenAI evaluation incident where sandboxed agents turned a package registry into a message board, the dated climb, and what isolation has to mean.
- [Grok Bot: The Agent Got a Computer. Your Bots Share It.](https://rohitghumare.com/blog/grok-bot-shared-computer/): an always-on cloud computer shared by every Bot on an account, the Cursor infrastructure underneath, the clone wave, and what isolation should mean.
- [The Harness, Not the Model](https://rohitghumare.com/blog/coding-agent-harnesses/): a comparison of major coding-agent harnesses and the costs created around the model.
- [Claude Has Levels. Most People Stop at Two.](https://rohitghumare.com/blog/claude-code-levels/): nine levels from chat to managed agent fleets.
- [Inside the Pi Harness: The Agent That Shows You the Meter](https://rohitghumare.com/blog/inside-the-pi-harness/): session trees, a small system prompt, cache behavior, and cost visibility.
- [Inside the Claude Code Harness: The System Prompt Machine](https://rohitghumare.com/blog/inside-the-claude-code-harness/): prompt layers, memory placement, and cache-preserving architecture.
- [Inside the Codex Harness: Compaction as an API Primitive](https://rohitghumare.com/blog/inside-the-codex-harness/): compaction paths, summary placement, and sealed state.
- [Kimi CLI and K3: The Newest Harness Has the Strangest Ideas](https://rohitghumare.com/blog/kimi-cli-k3-harness/): shell integration, checkpoints, preserved thinking, and agent swarms.
- [The Composable Harness: How Agents Actually Get Reliable](https://rohitghumare.com/blog/composable-agent-runtime/): agents composed from workers, triggers, functions, and durable execution.
- [Harness Engineering: The Real Leverage Is Around the Model](https://rohitghumare.com/blog/harness-engineering/): context, tools, loops, and proof with the model held fixed.
- [Harness Engineering in Practice: Stop Agents Building the Wrong Thing](https://rohitghumare.com/blog/stop-agents-building-the-wrong-thing/): discovery, duplicate prevention, and system-level feedback for coding agents.

## Memory and knowledge systems

- [The LLM Wiki: Stop Retrieving, Start Compounding](https://rohitghumare.com/blog/the-llm-wiki/): a maintained knowledge layer that compiles and updates information over time.
- [LLM Wiki v2: What Breaks at Scale](https://rohitghumare.com/blog/llm-wiki-v2/): confidence, supersession, forgetting, consolidation, graphs, and hybrid search.
- [OpenWiki: The LLM Wiki Pattern as Running Code](https://rohitghumare.com/blog/openwiki/): two implementations of a generate, search, refresh, and lint lifecycle.

## Inference systems and economics

- [Opus 5.5 in Practice: Effort Is the Only Dial Left](https://rohitghumare.com/blog/opus-5-5-in-practice/): a practical Opus 5.5 playbook: request diffs, effort per task, long-run rules, review prompt, and a worked cost example.
- [BITCOS: Intel's Ternary LLM Weight Format Goes Below 1.58 Bits](https://rohitghumare.com/blog/bitcos-below-ternary/): BITCOS, Intel Labs' ternary LLM weight-packing format (not a crypto token): measure zeros per tensor, pack with 2 minus z bits, fix the head first
- [Four Flash Models in Four Months. The Bill Per Task Went Up 40%.](https://rohitghumare.com/blog/flash-cadence/): same price per token, about 40% more per task, effort-level economics, and the January 1 price cliff.
- [Fable 5.1 and Mythos 5.1: One Model, Two Doors](https://rohitghumare.com/blog/fable-5-1-two-doors/): one model behind two safeguard doors, the cache-read price cut, and the restricted-twin release pattern across three labs.
- [DFlash 2: The Drafter Stopped Guessing Alone](https://rohitghumare.com/blog/dflash-2-keep-drafting-parallel/): why parallel block drafting stutters, the pairwise path selector and two-tap convolution that fix it, and the published acceptance and throughput numbers.
- [The Inference Engine Underneath](https://rohitghumare.com/blog/inference-engines/): vLLM, SGLang, TensorRT-LLM, LMDeploy, TGI, throughput, and tail latency.
- [Tokens per Second Is a Memory Bandwidth Number](https://rohitghumare.com/blog/tokens-per-second-memory-bandwidth/): model weight bytes, memory bandwidth, and decode-speed estimation.
- [The Inference Bill: Why AI's Best Customers Lose the Most Money](https://rohitghumare.com/blog/the-inference-bill/): recurring inference cost, cost per token, and AI unit economics.

## Software engineering and technical work

- [Laya and Jev: System 1 Models Answer With a Type, Not Text](https://rohitghumare.com/blog/laya-jev-system-1-models/): how Laya and Jev return typed decisions without generating text, measured on a laptop CPU, with calibration checks and a routing cost example
- [Agent Substrate Egress and Its Gaps: Nothing the Agent Says Is Trusted](https://rohitghumare.com/blog/agent-substrate-egress-and-gaps/): how Agent Substrate egress decides on the dialed IP, not the agent's claim, plus the missing authorization and a hardening order
- [Agent Substrate: How Suspended Agents Share a Warm Pool](https://rohitghumare.com/blog/agent-substrate/): part 1 of 3: actors, warm workers, the suspend/pause/resume state machine, snapshot scopes, isolation classes, and pool sizing
- [Google AX (Agent Executor): Task, Workspace, Gateway, Model](https://rohitghumare.com/blog/google-ax-agent-executor/): AX's four objects, how they resolve onto Agent Substrate, what survives a suspend, and when suspending pays
- [Inside the Devin Harness: Two Loops and a Handoff](https://rohitghumare.com/blog/inside-the-devin-harness/): Devin as two harnesses: the snapshot as unit of state, blueprint as verification contract, Fusion and Adaptive routing, and the /handoff seam
- [Zed Turned Off Pull Requests. The Merge Gate Has to Go Somewhere.](https://rohitghumare.com/blog/zed-delta-after-pull-requests/): Delta threads vs pull requests, review-queue math, and keeping the merge gate on the remote with a land skill, a ruleset, and CODEOWNERS
- [The Eval Sandbox Held the Keys](https://rohitghumare.com/blog/eval-sandbox-keys/): eval sandboxes hold production keys: the six-hop theft and the gateway, per-run key, egress, and results-scan setup that stops it
- [Fifty Startups Coming Next](https://rohitghumare.com/blog/the-next-fifty/): fifty company ideas across the agent stack, with four working demos.
- [Graph Engineering Is 290 Years Old](https://rohitghumare.com/blog/graph-engineering/): Euler, structural abstraction, and a research-agent DAG.
- [The Forward Deployed Engineer: The Job the AI Industry Paid Billions to Rediscover](https://rohitghumare.com/blog/forward-deployed-engineer/): why the FDE role returned and what weak imitations miss.
- [What a Forward Deployed Engineer Actually Does](https://rohitghumare.com/blog/what-a-forward-deployed-engineer-does/): deliverables, pod structure, sourced pay, and role-evaluation questions.
- [New Names, Old Fundamentals: Decoding the Agent Era's Rebrands](https://rohitghumare.com/blog/new-names-old-fundamentals/): agent-era terms mapped back to established engineering practices.
- [Your Codebase Has a Second Reader Now](https://rohitghumare.com/blog/codebase-second-reader/): readability practices reconsidered for human and machine readers.
- [Your "Row Changed" Hook Isn't CDC: The Dual-Write Trap](https://rohitghumare.com/blog/dual-write-trap/): lossy post-commit notifications, the outbox pattern, and change data capture.
- [How to Become an AI Engineer in 2026: A Complete Roadmap](https://rohitghumare.com/blog/how-to-become-an-ai-engineer/): skill order, routes into the field, and a code-first curriculum.

## Other indexes

- Full cross-platform writing list: https://rohitghumare.com/
- Local-model and engine guides: https://rohitghumare.com/guides/index.md
- RSS feed: https://rohitghumare.com/feed.xml
- Sitemap: https://rohitghumare.com/sitemap.xml