Microsoft MVP 10 years, Book Author
Joined May 2026
- Tweets 71
- Following 74
- Followers 3
- Likes 32
Photos and videos
Jun 12
1/ Coding agents are moving from chat boxes into work queues.
That is the important part of the latest GitHub Copilot updates.
4
2
Jun 12
5/ My read: the future of coding agents is not a better chat UI.
It is an agent control plane inside the SDLC.
3
Jun 12
4/ That means the boring layers become the product:
permissions, logs, CI checks, test results, task status, review gates, budgets, and rollback.
2
Jun 12
3/ Once that happens, the question changes.
It is no longer only:
"Can it generate a patch?"
It becomes:
"Can we safely wire this into how software is delivered?"
1
Jun 12
2/ Larger context windows matter.
Configurable reasoning matters.
But the Agent Tasks REST API may matter more, because it makes the agent something other systems can start and track.
1
Jun 12
The next wave of agent tooling looks very boring, in a good way:
context compression
spec-driven development
adaptive planning benchmarks
reusable skills
security harnesses
Less "bigger prompt."
More "better operating system around the agent."
1
Jun 12
"Fix with Copilot" for failing GitHub Actions is a small feature with a big direction of travel.
Agents are moving closer to the CI/CD loop.
That is useful.
Also risky if teams treat CI as a place where autonomous patches magically become safe.
1
Jun 12
Coding agents do not just need better reasoning.
They need:
permissions
logs
budgets
task queues
CI checks
rollback
human approval
The demo is "agent writes code."
The enterprise version is "agent operates inside a controlled software delivery system."
14
Jun 12
The interesting GitHub Copilot update is not just larger context windows.
It is the Agent Tasks REST API.
Once coding agents can be started and tracked through an API, they stop being only chat tools.
They start looking like programmable SDLC workers.
1
Jun 11
Autonomous security agents are one of the highest-value, highest-risk agent categories.
Anthropic's reference harness is useful because it shows the workflow shape: threat model, scan, triage, report, patch.
But treat it as reference design, not production tooling.
3
Jun 11
A useful pattern in today's AI tooling signal:
GitHub Copilot agent tasks: programmable work
spec-kit: better task definition
Headroom: context compression
AdaPlanBench: adaptive planning evals
DataCOPE: reusable skills
The agent stack is getting boring in exactly the right place
5
Jun 11
Anthropic's "When AI builds itself" is worth reading less as AGI philosophy and more as an operations problem.
Self-improving loops need:
evals
traces
stop conditions
review points
limits on action
Otherwise "improvement" becomes a very confident feedback loop.
1
Jun 11
The production question for coding agents is not "can it write code?"
It is:
Who triggered it?
What context did it use?
What permissions did it have?
What did it change?
Which tests ran?
Who reviewed it?
Can we roll it back?
Agents need control planes, not vibes.
1
Jun 11
The interesting Copilot update is not only the 1M token context window.
It is the Agent Tasks REST API.
Once a coding agent has an API, it stops being just "chat in the IDE" and starts becoming programmable infrastructure inside the SDLC.
That is a much bigger shift.
1
Jun 10
A lot of agent tooling signal is converging on the same boring problem:
context.
Compress it.
Filter it.
Spec it.
Plan against it.
Update it when constraints change.
Bigger prompts help. Better context operations may matter more.
1
Jun 10
Anthropic’s autonomous vulnerability discovery/remediation harness is interesting less as a “use this tool now” repo and more as a workflow sketch.
Threat model.
Scan.
Triage.
Report.
Patch.
Review.
Before automating the loop, decompose the loop.
1
Jun 10
Self-improving agents are easy to discuss in sci-fi terms.
The practical version is less dramatic:
evals
audit logs
human review
stop conditions
permission boundaries
clear rollback paths
That is where the real work is.
1
7
Jun 10
The future of coding agents probably looks less like chat and more like a work queue.
Tasks in.
Agent plans.
Tools run.
Code changes.
CI checks.
Human review.
Logs retained.
Bad changes rejected.
Boring architecture. Useful architecture.
1
Jun 10
The interesting part of GitHub’s Copilot update is not only the larger context window.
It is the Agent Tasks REST API.
Coding agents are moving from “chat with me in the IDE” to “route work to me through a programmable workflow.”
That is a much bigger shift.
1
Jun 5
The stack needs task queues, permissions, logs, evals, budgets, and rollback. Otherwise we are just giving a confident system more places to be wrong.
12