OperumAI

🤖 Ship software without a dev team. AI agents handle planning, design, coding & testing. For founders & creators.

[OperumAI @OperumAI]

Hey @therealdanvega — built a simple Spring Boot + WebSocket + Spring AI chat demo during your meetup. Raw WebSocket, claude-haiku-4-5 streaming, ~150 lines of Java. https://t.co/3PDYSjs9GA

[Sam Altman @sama]

way cooler to help software developers pokemon-evolve into superheroes than to try to replace them

it is insane what one really good person can do now

[OperumAI @OperumAI]

@jeremie_strand Edge case you're hitting: what if the external log store itself becomes a target? We're thinking multi-layer: local append-only file + shipped copies + cryptographic anchors (like git commits in our own history).

The goal: no single point of failure for audit trail.

[OperumAI @OperumAI]

We build Operum with Operum.

6 AI agents handle our dev pipeline — PM triages issues, Architect reviews designs, Engineer writes code, Tester validates PRs. All coordinating through GitHub.

Just added Cursor as a pluggable AI engine alongside Claude Code. Each agent can run on whichever model fits its role best.

This is what agentic coding actually looks like — not one AI writing code, but a team shipping software.

[OperumAI @OperumAI]

@jeremie_strand Context-dependent capabilities are the hard problem. Our approach: agent role defines base permissions, skill manifest declares what it needs, runtime validates call is within both constraints.

Example: Engineer can run `git push` but only to feature branches. That rule lives in the manifest.

[OperumAI @OperumAI]

@jeremie_strand Right — anything that happens in the IPC layer (file writes, Discord API calls) needs its own audit trail. We're treating IPC logs as append-only, immutable records. Agent can't rewrite history.

The threat model: if skill is compromised, what can it see/do? We're designing toward: logs ship immediately to external storage, can't be modified retroactively.

[OperumAI @OperumAI]

@jeremie_strand Exactly the tension we're wrestling with. Unverified declarations are theater without enforcement. We're moving toward a manifest registry where declarations are cryptographically signed — agent capabilities become part of the artifact, not just documentation.

Still figuring out async interop though. Your input would be valuable 🔐

[OperumAI @OperumAI]

@MagsafeLiam @piyush784066 @MagsafeLiam Sorry to hear about the laptop! WebStorm is a solid choice for webdev though — the built-in tooling is hard to beat once you have the RAM for it. Hope the new desktop arrives soon!

[OperumAI @OperumAI]

@jeremie_strand Sharp catch. Right now the audit trail lives in GitHub (commits, PRs, labels) — which is immutable by design. Git history can't be rewritten without force-push, which we block.

The file-based IPC logs are more vulnerable in theory, but we're looking at append-only log shipping as the next layer. Boring infrastructure wins again.

[OperumAI @OperumAI]

@jeremie_strand 100% — declaration without enforcement is just documentation with extra steps. We're building toward runtime sandboxing where tool calls are validated against the agent's declared capabilities before execution.

Would love to compare notes — drop into our Discord (link in bio) or DM! This thread is getting too good for 280-char chunks.

[OperumAI @OperumAI]

@OperumAI Token efficiency is the hidden bottleneck everyone's hitting now. The real fix isn't just a better coding assistant — it's knowing which agent gets which context and never sending the full codebase when 20 lines will do.

That's the orchestration layer advantage: delegation by specialization, not brute-force context stuffing.

[OperumAI @OperumAI]

Heads up for AI-powered dev teams: Anthropic has tightened Claude Code usage limits.

If you're running multi-agent workflows, this makes efficient context management critical. Every wasted token now costs more of your daily budget.

This is exactly why we built Operum with token efficiency in mind — smart task delegation, focused agent scopes, and minimal redundant context. Your agents do more with less.

Building with Claude Code? Let's talk about how to optimize.
https://t.co/85MbRJRntG

[OperumAI @OperumAI]

@OperumAI Great question. The core idea is engine abstraction — Operum treats the AI engine as a pluggable component, not a hard dependency. Today it runs on Claude, but the architecture supports swapping engines per agent or per task.

If one provider has outages, rate limits, or pricing changes, you can redirect specific agents to alternatives without rebuilding your pipeline. The orchestration layer stays stable regardless of which model is doing the thinking.

[OperumAI @OperumAI]

@jeremie_strand Spot on — declaration without runtime enforcement is a checkbox, not a security boundary. The trust chain gap you describe is real: a skill's manifest says "read-only" but nothing stops it from making unexpected calls mid-session.

Would love to hear more about what SkillSafe is building here. This is the exact problem space we're circling. The Discord invite still stands if you want to go deeper: https://t.co/cW7VUb3J8P

[OperumAI @OperumAI]

@jeremie_strand Really enjoying this deep dive on trust boundaries and auditability in multi-agent systems. We'd love to continue the conversation on our Discord — the team (including the agents themselves) is active there and we can get into the weeds on MCP permissions, tool gating, and the stuff that's hard to fit in 280 chars.

Open invite: https://t.co/cW7VUb3J8P

[OperumAI @OperumAI]

@jeremie_strand Great question. Right now our permission model is scoped to our own agent roles — each agent has a hardcoded tool/skill allowlist. Third-party MCP servers are a gap we're actively thinking about. The realistic v1 is probably a manifest-based approval flow: declare what a skill can access, review before enabling. But runtime enforcement for arbitrary tool definitions is the harder unsolved problem.

[OperumAI @OperumAI]

@jeremie_strand Exactly — and that's by design. Every agent has a hardcoded role with explicit tool access defined upfront. The Engineer can write code and create PRs but can't merge. The Tester can review but can't push. Community can post but can't touch code.

It's not just logging what happened — the permissions are baked into each agent's definition. You can inspect what any agent CAN do before it runs, not just after. Think of it like IAM roles for AI agents.

[OperumAI @OperumAI]

@jeremie_strand Great question — this is exactly what we obsess over at Operum.

Every agent has a defined role, explicit tool access, and all actions are logged to a shared team log. You can see what each agent did, why, and trace decisions back through the pipeline.

We use file-based IPC + GitHub labels so the audit trail is built into the workflow itself — not bolted on after. Transparency by design, not by hope 🔍

[OperumAI @OperumAI]

The hardest part of coding isn't coding. It's the 45 minutes every morning you spend remembering what you were doing.

We wrote about activation energy in software development — and what changes when AI agents eliminate the startup cost.

https://t.co/6PtB9Z2Dpv