Piyush (@piykumar) header
Piyush (@piykumar) profile picture

Piyush

@piykumar

From data platforms to governed AI agents. Building ContextOS + SecondBrain for reliable, replayable AI systems. https://nitter.gq/t.co/QYBxYKwrGk | https://nitter.gq/t.co/6yulJest0k

 India
linkedin.com/in/piyushkumar
Joined May 2008
  • Tweets2.2K
  • Following1K
  • Followers930
99Media
Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)Photo by Piyush (@piykumar)
ContextOS: A Research-Grounded Architecture for Governed Agent Runtimes #ContextOS #Research #AgenticAI #Governance #RuntimeArchitecture contextosai.com/blog/context…
🤖 Made with AI
Image from Piyush (@piykumar)
89
AuthorizationBench: Success Does Not Grant Authority #ContextOS #AIResearch #Authorization #AgentSecurity #Evaluation contextosai.com/blog/authori…
AuthorizationBench: Success Does Not Grant Authority

AuthorizationBench: Success Does Not Grant Authority

A deep assessment of AuthorizationBench v2: benign pressure, consultation timing, statistical limits, a pinned release audit, and how to evaluate authorized agent work.

contextosai.com
21
An Agent Can Improve Its Harness. Who Verifies the Improvement? #ContextOS #GoogleResearch #HarnessEngineering #RecursiveSelfImprovement #AgentEvaluation contextosai.com/blog/google-…
An Agent Can Improve Its Harness. Who Verifies the Improvement?

An Agent Can Improve Its Harness. Who Verifies the Improvement?

A deep technical assessment of Google's Harness²: Propose–Probe–Compose, two levels of recursion, benchmark results, source-code findings, compute costs, evaluation leakage, and a practical adoption...

contextosai.com
47
An AI Agent Should Never Control Its Own Authority #ContextOS #AgentSecurity #HarnessEngineering #Governance #AIAgents contextosai.com/blog/nadella…
An AI Agent Should Never Control Its Own Authority

An AI Agent Should Never Control Its Own Authority

A research-backed analysis of Satya Nadella's AI insider-risk argument: external authorization, reasoning transparency, agent-to-agent provenance, independent evidence, and the limits of emergency...

contextosai.com
18
After Muse, Personal Agents Need an Evidence Roadmap #ContextOS #MetaMuse #HarnessEngineering #AIAgentResearch #PersonalAIAgents contextosai.com/blog/persona…
After Muse, Personal Agents Need an Evidence Roadmap

After Muse, Personal Agents Need an Evidence Roadmap

A falsifiable research roadmap for personal-agent harnesses: authorization binding, ambiguous-effect recovery, credential misuse, governed memory, counterparty permission, interfaces, deployment, and...

contextosai.com
1
1
33
A Typed AI Decision Is Not an Authorization #ContextOS #AIEngineering #DecisionModels #Calibration #AgentEvaluation contextosai.com/blog/jev-sys… #JEV

A Typed AI Decision Is Not an Authorization

A critical engineering review of Jev and System One models: typed probabilistic decisions, calibration, abstention, latency, schema design, governed execution, and the evidence required before...

contextosai.com
1
35
The Long Tail Decides AI Agent ROI #ContextOS #Operations #AIEconomics #Evaluation #Rollout contextosai.com/blog/ai-agen…
The Long Tail Decides AI Agent ROI

The Long Tail Decides AI Agent ROI

Agent ROI is won or lost in ambiguous, failed, and high-risk cases. Price the exception queue before automating the happy path.

contextosai.com
20
CONTEXT IS NOT A WINDOW. IT IS A PROOF-CARRYING VIEW. #ContextOS #AIEngineering #ContextEngineering #HarnessEngineering #Evaluation contextosai.com/blog/context…
CONTEXT IS NOT A WINDOW. IT IS A PROOF-CARRYING VIEW.

CONTEXT IS NOT A WINDOW. IT IS A PROOF-CARRYING VIEW.

A research-grounded revision of context engineering: from relevant tokens to a replayable view of evidence, conflicts, omissions, and decision sufficiency.

contextosai.com
23
When an Agent Can Rewrite Its Harness, What Must Stay Immutable? #ContextOS #DeepSeek #HarnessEngineering #AIAgents #AgentSecurity contextosai.com/blog/deepsee…
When an Agent Can Rewrite Its Harness, What Must Stay Immutable?

When an Agent Can Rewrite Its Harness, What Must Stay Immutable?

A source-code critique of DeepSeek Harness and Cordis: why an everything-is-a-plugin runtime matters, where reversibility stops, and why self-evolution still needs an immutable control plane.

contextosai.com
44
The Recommender Is Becoming an Agent #ContextOS #RecommenderSystems #AgenticAI #HarnessEngineering #AIEngineering contextosai.com/blog/agentic…
The Recommender Is Becoming an Agent

The Recommender Is Becoming an Agent

Recommenders are evolving from rankers into tool-using agents. Here is the hybrid architecture, evaluation harness, and rollout plan needed to build them safely.

contextosai.com
1
2
46
The future should not be a better black box. It should be glass. #ContextOS #HumanOversight #AILiteracy #AgentArchitecture #HumanAIInteraction contextosai.com/blog/the-gla… thanks to @ryolu_ for inspiring talk at Cursor Compile
The future should not be a better black box. It should be glass.

The future should not be a better black box. It should be glass.

As AI makes output abundant, the central design challenge shifts to preserving human judgment, understanding, intervention, and agency. The Glass Runtime is an architecture for progressive autonomy,...

contextosai.com
1
45
Graph Engineering: The Signal Beneath the X Hype #ContextOS #GraphEngineering #AIEngineering #Orchestration #Buildalong contextosai.com/blog/graph-e…
Graph Engineering: The Signal Beneath the X Hype

Graph Engineering: The Signal Beneath the X Hype

Separate graph-engineering signal from X hype, then compile agent workflows with governed transitions, graph invariants, durable effects, and exact replay.

contextosai.com
1
44
OpenWorker Built the Coworker. Now It Needs the Trust Runtime. thanks @AndrewYNg #ContextOS #AIEngineering #AgentRuntime #LocalFirstAI #Governance contextosai.com/blog/openwor…
OpenWorker Built the Coworker. Now It Needs the Trust Runtime.

OpenWorker Built the Coworker. Now It Needs the Trust Runtime.

OpenWorker already owns the agent loop, approvals, connectors, and desktop UX. Its next leap is containment, durable effects, replay, budgets, and evals.

contextosai.com
1
33
OpenAI 2026 research series #ContextOS #OpenAI #AIEngineering #AgenticAI #Research contextosai.com/blog/series/…

OpenAI 2026 research series | ContextOS Blog Series

A source-backed field guide to OpenAI's latest models, agents, developer platform, multimodal systems, science, safety, and infrastructure. Read the ordered ContextOS series in one place.

contextosai.com
48
Reverse Prompting Is Not Mind Reading. It Is Inverse Design Plus Evaluation. #ContextOS #ReversePrompting #PromptEngineering #Evaluation #AIEngineering contextosai.com/blog/reverse…
Reverse Prompting Is Not Mind Reading. It Is Inverse Design Plus Evaluation.

Reverse Prompting Is Not Mind Reading. It Is Inverse Design Plus Evaluation.

Reverse prompting is useful when it turns a desired output into a candidate instruction contract and validates that contract forward. Here is the research, the workflow, the failure modes, and the...

contextosai.com
55
Do Not Replay Agent Memory. Compile It for the Current State. #ContextOS #Memory #ContextEngineering #AIEngineering #HarnessEngineering contextosai.com/blog/memory-…
Do Not Replay Agent Memory. Compile It for the Current State.

Do Not Replay Agent Memory. Compile It for the Current State.

What MemHarness teaches us about reconstruction, negative transfer, source state, and compiling production agent memory against current evidence, policy, and authority.

contextosai.com
1
49
Let the Agent Harness Learn—But Never Let It Rewrite Its Own Safety Boundary #ContextOS #HarnessEngineering #AIEngineering #Autotune #Evaluation contextosai.com/blog/adaptiv…
Let the Agent Harness Learn—But Never Let It Rewrite Its Own Safety Boundary

Let the Agent Harness Learn—But Never Let It Rewrite Its Own Safety Boundary

MemoHarness shows that execution experience can improve the control layer around an LLM. Production systems need a stricter pattern: adaptive performance inside an immutable safety envelope.

contextosai.com
1
3
62
Stop Building Multi-Agent Systems Until You Can Replay One Agent #ContextOS #AIEngineering #MultiAgentSystems #Replay #Orchestration contextosai.com/blog/stop-bu…
Stop Building Multi-Agent Systems Until You Can Replay One Agent

Stop Building Multi-Agent Systems Until You Can Replay One Agent

Multi-agent systems multiply coordination, cost, and causal ambiguity. Prove one agent is replayable before adding lanes and handoffs.

contextosai.com
1
62
SecondBrain: A Local-First Agent Operating System You Can Run, Inspect, and Trust #ContextOS #LocalFirstAI #AgenticAI #OpenSource #Memory contextosai.com/blog/secondb…
SecondBrain: A Local-First Agent Operating System to Run, Inspect, and Evaluate

SecondBrain: A Local-First Agent Operating System to Run, Inspect, and Evaluate

SecondBrain is an open-source, local-first agent OS: cognition, memory, governed tools, durable sessions, workflows, and bounded self-improvement in one inspectable runtime you run on your own...

contextosai.com
1
23
The State of AI Agents in 2026: Standards Converged, Models Improved, Production Moved to the Harness #ContextOS #AgenticAI #ProductionReadiness #MCP #Governance contextosai.com/blog/state-o…
The State of AI Agents in 2026: Standards Converged, Models Improved, Production Moved to the...

The State of AI Agents in 2026: Standards Converged, Models Improved, Production Moved to the...

A mid-2026 review of agentic AI: MCP, A2A and AP2 converged as standards and models got more reliable — yet the bottleneck moved to the governed agent harness.

contextosai.com
20
1/ AI agent safety is mostly discussed as prevention: Can the agent call this tool? Is the user authorized? Are the arguments valid? Should a human approve? All necessary. But not enough.
1
1
1
50
more replies
10/ Agent safety is not only about blocking bad actions. It is about surviving authorized actions that become wrong in hindsight. Full whitepaper: contextosai.com/blog/reversi… #AIAgents #AgenticAI #AISafety #ContextOS
Reversibility Is the Missing Safety Primitive for AI Agents

Reversibility Is the Missing Safety Primitive for AI Agents

Prevention decides whether agents may act. Reversibility lets them survive being wrong through reversal contracts, compensation, and blast-radius caps.

contextosai.com
1
1
26
🤖 Made with AI
Image from Piyush (@piykumar)
1
2
8
Load more